A moving target tracking method combining active and passive information fusion and ensemble learning

Abstract

The present invention discloses a method for tracking a moving target that combines active and passive information fusion with integrated learning, which mainly solves the problems of the existing technology that a single data source is easily affected by noise and outliers, the model is highly complex and has poor stability. The implementation steps are: fusing active detection data and passive detection data to obtain observation data; using polynomial fitting least squares method and extreme learning machine to fit the observation data to obtain basic learner 1 and basic learner 2 respectively; combining the prediction results of the two learners through the stacking method of integrated learning to obtain the predicted state of the target; performing Gaussian filtering and smoothing on the predicted state of the target; calculating the movement speed of the target; judging whether it is the last set of observation data, and if so, ending the tracking. The present invention improves the stability of the model and achieves efficient and accurate tracking of the target by combining active and passive information fusion with integrated learning methods.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning technology, and further relates to a moving target tracking method that combines active and passive information fusion with ensemble learning in the field of target tracking technology. The present invention can be used to track a moving target by solving the motion elements of the moving target. Background Art

[0002] Target motion analysis is a method for estimating motion parameters of a moving target, such as range, orientation, and velocity, from sensor data. Major target motion analysis methods include passive and active methods. Passive target motion analysis uses only the target's orientation angle to estimate and track motion parameters using geometric models and algorithms such as least squares and Kalman filtering. Active target motion analysis enhances prediction capabilities by providing additional information, such as distance. Among traditional target motion analysis algorithms, the least squares method is suitable for linear problems and has low model and algorithmic complexity. However, it is sensitive to noise, suffers from poor real-time performance, and lacks adaptability to diverse situations. With the development of deep learning, the application of neural networks in target motion analysis has steadily increased. Neural networks can predict target trajectories based on historical data, offering good real-time performance and accuracy, and are applicable to both linear and nonlinear problems. Extreme learning machines (ELMs) have attracted particular attention due to their rapid training and strong generalization capabilities.

[0003] The 20th Institute of China Electronics Technology Group Corporation disclosed a method for tracking moving targets that integrates Kalman filtering and DDQN algorithms in its patent application document "A method for tracking moving targets of unmanned aerial vehicles that integrates Kalman filtering and DDQN algorithms" (application number 202011440212.3, application publication number CN112435275A). The implementation steps of this method are to accurately estimate the target motion state to obtain the target position and velocity, and then combine the drone's own state information as the neural network input, use the drone's acceleration and angular velocity as the action output, learn through the DDQN algorithm, complete the training of the flight strategy network, and realize the drone's autonomous tracking decision for the maneuvering target. Although this method improves the error of direct ranging using sensors in traditional unmanned aerial vehicle target tracking tasks, the method still has the disadvantages of high model complexity and failure to effectively integrate the advantages of different data sources, resulting in low tracking accuracy and stability in the case of large noise or complex environment.

[0004] Chongqing University, in its patent application, "A Method and Apparatus for Multi-Sensor Information Fusion" (Application No. 202311121629.7, Publication No. CN 117113276 A), discloses a moving target tracking method using multi-sensor information fusion. This method involves performing a spatiotemporal preprocessing algorithm based on production rules on sensor layer data to obtain preprocessed data; wherein the sensor layer data is acquired from multiple sensors; performing feature-level fusion on the preprocessed data using a neural network multi-classifier to obtain feature-fusion data; performing decision-level fusion on the feature-fusion data using a support-corrected iterative fusion method to obtain decision-level data; and uploading the decision-level data for decision-making. However, this method suffers from the following drawbacks: the complex neural network multi-classifier and support-corrected iterative method are used for multi-level feature fusion and decision fusion, resulting in high model complexity, high computational resource consumption, and time-consuming training and inference processes. Summary of the Invention

[0005] The purpose of the present invention is to address the shortcomings of the above-mentioned existing technologies and propose a moving target tracking method that combines active and passive information fusion with integrated learning. The method aims to solve the many limitations faced by existing target motion analysis technologies when dealing with complex environments, including the influence of noise and outliers caused by a single data source, high model complexity and poor model stability, which make it difficult to meet the actual needs of tracking moving targets.

[0006] The specific idea for achieving the purpose of the present invention is to use the least squares method and the extreme learning machine as basic learners to perform target tracking prediction. The least squares method solves the model parameters by minimizing the sum of squared errors between the predicted value and the true value, and the extreme learning machine is trained by randomly selecting the input layer weights and optimizing the output layer weights, avoiding the complex calculations of the back propagation algorithm in the traditional neural network. In this way, model training and prediction can be completed quickly. In order to further improve the performance, the present invention adopts an integrated learning method, which combines the linear fitting ability of the least squares method with the nonlinear fitting ability of the extreme learning machine through the stacking method to give full play to the advantages of both. At the same time, combined with active and passive information fusion technology, only passive detection is used outside the active detection range, and passive detection is used within the active detection range to correct and supplement abnormal values, thereby enhancing the stability of target tracking.

[0007] To achieve the above object, the specific steps of the present invention are as follows:

[0008] Step 1: Fuse the active and passive detection data of the target at the current moment:

[0009] Step 1.1: Determine whether the distance between the target and the detection source is greater than or equal to a threshold. If so, use only passive detection data to predict the target's azimuth at the current moment. Otherwise, proceed to step 1.2. The threshold is equal to the maximum distance at which the active detector can effectively capture the target.

[0010] In step 1.2, if the difference between the azimuth of the current active detection data and the azimuth of the previous active detection data exceeds the dynamic threshold at the current moment, the active detection data is determined to be abnormal. The azimuth of the passive detection data is used to perform a fusion correction on the azimuth of the active detection data through weighted fusion.

[0011] Step 1.3: Using the fused and corrected azimuth detection values, calculate the target's observed position on the x-axis and y-axis through vector decomposition, and generate the target's observation information at the current moment. This observation information includes the target's two-dimensional position, distance, and azimuth.

[0012] Step 2: Use the polynomial fitting least squares method to fit the observed data of the target at the current moment to obtain the basic learner 1;

[0013] Step 3: Use the extreme learning machine to fit the observation data of the target at the current moment to obtain the basic learner 2;

[0014] Step 4: Use the stacking method in ensemble learning to combine the prediction results of base learner 1 and base learner 2 to obtain the prediction tracking result of the target at the current moment, that is, the three elements of the target prediction state and

[0015] Step 5: Prediction of the azimuth angle element based on the target prediction state at the current moment The prediction of the target azimuth angle element at the previous moment Calculating the difference between the two azimuth angle predictions. If the difference is greater than a prediction threshold, it is determined that the target has turned at the current moment; the prediction threshold is equal to the target's minimum turning angular velocity multiplied by the sensor's sampling time interval;

[0016] Step 6: Use Gaussian filter to predict the three elements of the target state and Perform smoothing;

[0017] Step 7, calculate the target's movement speed at the current moment;

[0018] Step 8: Determine whether the current observation data is the last set of observation data. If so, execute step 9; otherwise, execute step 2;

[0019] Step 9: End tracking of the target.

[0020] Compared with the existing technology, the present invention has the following advantages:

[0021] First, since the present invention uses the least squares method and the extreme learning machine as basic learners, it can perform rapid training and obtain tracking prediction results, overcoming the problem of high model complexity of existing methods, so that the present invention improves the tracking efficiency of moving targets.

[0022] Second, since the present invention uses an integrated learning machine learning method, the linear fitting ability of the least squares method is combined with the nonlinear fitting ability of the extreme learning machine through the stacking method, the advantages of the two models are complemented, and the problem of poor model stability of the existing method is overcome. The present invention improves the stability of the model and at the same time improves the accuracy of the estimation of the moving target state.

[0023] Third, since the present invention adopts active and passive information fusion technology, passive detection data is used as a supplement to active detection data. By only using passive detection outside the range of active detection, and using passive detection as an outlier correction and supplement to active detection within the range of active detection, the noise and outlier effects caused by the single data source in the existing method are overcome, so that the present invention can achieve more accurate moving target tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flow chart of the present invention;

[0025] Figure 2 、 Figure 3 、 Figure 4 、 Figure 5 Graphs showing simulation experiment results for different scenarios of the present invention. DETAILED DESCRIPTION

[0026] The implementation steps and simulation effects of the embodiments of the present invention are further described below with reference to the accompanying drawings.

[0027] Reference Figure 1 , further describing the implementation steps of the embodiment of the present invention.

[0028] Step 1: Target motion state modeling.

[0029] In the embodiment of the present invention, the motion state of the target is first modeled. Assuming that the target motion occurs in a two-dimensional plane coordinate system, the motion state at each moment can be represented by a state vector s containing eight components. t Indicates that:

[0030]

[0031] Then, the target's motion state is updated according to the following state transition equation:

[0032]

[0033] Thus we get

[0034] Among them, s t represents the motion state of the target at time t, x t 、y t Respectively represent the position of the target on the x-axis and y-axis at time t, They represent the speed of the target on the x-axis and y-axis at time t, respectively, t Indicates the distance between the target and the origin of the coordinate axis at time t, θ t Indicates the azimuth angle of the target relative to the origin of the coordinate axis at time t. t+1 represents the motion state of the target at time t+1, x t+1 、y t+1 Respectively represent the position of the target on the x-axis and y-axis at time t+1, They represent the speed of the target on the x-axis and y-axis at time t+1, respectively. t+1 Indicates the distance between the target and the origin of the coordinate axis at time t+1, θ t+1 It represents the azimuth angle of the target relative to the origin of the coordinate axis at time t+1, and T represents the sampling interval.

[0035] During the entire motion of the target, the acceleration a of the target on the x-axis and y-axis x and a y The size remains unchanged.

[0036] Set the target motion mode and maneuvering mode. Typical motion modes include constant speed, and maneuvering modes include variable speed.

[0037] If the target moves at a constant speed, the acceleration of the target on the x-axis and y-axis is x and a y Set to 0; if the target moves at a variable speed, a x and a y At least one of them is not 0.

[0038] Step 2: Fuse the active and passive detection data of the target at the current moment.

[0039] The active detection and passive detection data of the target refer to that the active detection data is the target azimuth and distance information detected by the active detection sensor, and the passive detection data is the target azimuth information detected by the passive detection sensor.

[0040] The data used in the embodiments of the present invention includes active detection data and passive detection data passive=[θ t passive ].

[0041] in, represents the active detection distance of the target at the tth moment, θ t active represents the active detection azimuth of the t-th target; θ t passive represents the passive detection azimuth of the target at the tth moment.

[0042] The information fusion scheme is based on the distance of the target and is divided into the following two cases:

[0043] When the distance of the target is greater than the distance threshold D, only the passive detection data θ t passive , as the azimuth angle θ of the target motion state at this time t The prediction of the distance threshold D is equal to the maximum distance at which the active detector can effectively capture the target. In the embodiment of the present invention, D is set to 2000m.

[0044] When the distance to the target is less than the distance threshold D, if the difference between the azimuth of the current active detection data and the azimuth of the previous active detection data exceeds the dynamic threshold at the current moment, the active detection data is judged to be abnormal. The dynamic threshold is determined by the following formula:

[0045]

[0046] Among them, ε t represents the dynamic threshold at the tth moment, d t Indicates the distance between the target and the detection source at the tth moment, d max Indicates the maximum distance that the active detector can actively detect, △θ max Indicates the maximum detection error of the active detector, and its value is determined by the calibration data of the detector. max =2°.

[0047] For outlier correction, the azimuth of the passive detection data is used to perform fusion correction on the azimuth of the active detection data. A weighted fusion method can be used to combine the active and passive detection data to obtain a more accurate azimuth detection value:

[0048]

[0049] in, represents the target azimuth after fusion correction at the tth moment, θ t active represents the azimuth of active detection at the tth moment, θ tpassive represents the azimuth of passive detection at the t-th moment, and α represents the weight of active detection data in the fusion correction, and its value range is 0≤α≤1. In the embodiment of the present invention, α=0.7.

[0050] After obtaining the fused azimuth detection value Then, the vector decomposition is performed according to the following formula to calculate the target's observation position on the x-axis and y-axis, and generate the target's observation data at the current time t The observation information includes the target's two-dimensional position, distance and azimuth:

[0051]

[0052] in, They represent the observed positions of the target on the x-axis and y-axis at the t-th moment respectively. represents the active observation distance of the target at the tth moment, Represents the target azimuth after fusion correction at the tth moment.

[0053] Step 3: Normalize the observed data.

[0054] Obtaining observation data And the time series t=[t1,t2,...,t n ], n represents the total number of observation data determined by the requirements of solution time and solution accuracy. In the embodiment of the present invention, n=80. In order to perform subsequent tracking and prediction, the data needs to be normalized first. According to the following formula, o t and t are normalized:

[0055]

[0056] in, and Represents the normalized observation data and time series data respectively. Divisor vector s o and s t is the maximum value of each observation dimension and time series dimension, that is: s t =max(t).

[0057] Step 4: Use the polynomial fitting least squares method to fit the observed data of the target at the current moment.

[0058] Based on the normalized observation data, the target's motion trajectory is fitted using the polynomial fitting least squares method. Specifically, the three state elements of the normalized moving target are: and Perform fitting and obtain their fitting functions f with respect to time tx (t), f y (t) and f θ (t), the fitting steps are as follows:

[0059] The first step is to construct the matrix A as follows:

[0060]

[0061] Where A represents an n×(m+1) matrix, n represents the total number of observation data determined by the requirements of solution time and solution accuracy, m represents the highest order of polynomial fitting, t1, t2, t n Respectively represent the t1th moment, t2th moment, t n At each moment, calculate the elements of A;

[0062] The second step is to calculate the unknown coefficients a, b, and c according to the following formula:

[0063] a=(A T A) -1 A T x

[0064] b=(A T A) -1 A T y

[0065] c=(A T A) -1 A T θ

[0066] Among them, the superscript T represents the transpose operation, and the superscript -1 represents the inverse operation. Respectively represent the t1th moment, t2th moment, t n The observed position of the target on the x-axis at each moment; Respectively represent the t1th moment, t2th moment, t n The observed position of the target on the y-axis at each moment; Respectively represent the t1th moment, t2th moment, t n The target azimuth after fusion correction at each moment;

[0067] The third step is to get f x (t), f y (t), f θ (t) Three polynomial fitting functions form the basic learner 1:

[0068]

[0069] Among them, f x (t), fy (t), f θ (t) represents the three state elements of the normalized target at the tth moment and The polynomial fitting function, a k 、b k 、c k They represent the undetermined coefficients of the kth polynomial order in the polynomial fitting function, and their value range is k=0,1,2,...,m.

[0070] Step 5: Use the extreme learning machine to fit the observed data of the target at the current moment.

[0071] Based on the normalized observation data, the steps for using the extreme learning machine to fit the observation data of the target at the current moment are as follows:

[0072] The first step is to determine the relevant parameters of the extreme learning machine network structure, including the dimension of the input layer, the number of hidden layer neurons, and the type of activation function;

[0073] The second step is to randomly initialize the hidden layer weights and bias parameters of the extreme learning machine network;

[0074] The third step is to calculate the element h of the hidden layer output matrix H according to the following formula. ij :

[0075]

[0076] Among them, h ij Represents the element in row i and column j of the hidden layer output matrix H, t p Indicates the tth p At each moment, 1≤p≤n, the value of i is equal to the value of p, c q The center of the qth neuron can be determined by random selection method, randomly selected from t p Select a moment as c q The value of σ q represents the width parameter of the qth neuron, σ q =σ, σ is the average distance between all input samples,

[0077] The fourth step is to solve the hidden layer output weight matrix β according to the formula:

[0078] β=(H T H+λI) -1 H T z

[0079] Among them, λ represents the regularization term to prevent the matrix from being irreversible, Cond represents the condition number of the matrix; z represents the observed data of the target, respectively. and Substituting in, I is the identity matrix;

[0080] The fifth step is to train the extreme learning machine to obtain and About the extreme learning machine fitting function h at the tth moment x (t), h y (t) and h θ (t), and use it as the basic learner 2 to participate in the ensemble learning. Their fitting function form is:

[0081]

[0082] Among them, β x,q , β y,q , β θ,q Respectively represent the output weight matrix β and and The corresponding weight; h q (t) represents the output of the qth neuron in the hidden layer.

[0083] Step 6: Use ensemble learning to obtain the target prediction and tracking results at the current moment.

[0084] The stacking method in ensemble learning is used to combine the prediction results of basic learner 1 and basic learner 2 to obtain the prediction tracking result of the target at the current moment, that is, the three elements of the target prediction state and The specific steps are as follows:

[0085] In the first step, a linear regression model is selected as the meta-learner g(t). The coefficient w of the meta-learner is solved by minimizing the sum of squared errors. The form of the meta-learner is:

[0086] g(t)=w0+w1f(t)+w2h(t),

[0087] Where w = (w0, w1, w2) T represents the coefficient of the meta-learner, w=(X T X) -1 X T z, z represent the observation data of the target, respectively and Substitute, f(t), h(t) represent the prediction results of base learner 1 and base learner 2 at the tth moment, respectively. For the target state element and Train a meta-learner g separately x (t), g y(t) and g θ (t), that is, for the target state element Use f x (t) and h x (t) is the base learner of the meta-learner, for the target state element Use f y (t) and h y (t) is the base learner of the meta-learner, for the target state element Use f θ (t) and h θ (t) serves as the base learner of the meta-learner;

[0088] The second step is to construct a stacked feature matrix X, which contains the prediction results of the base learner 1f(t) and the base learner 2h(t) and the constant term:

[0089]

[0090] Among them, f(t1), f(t2), f(t n ) respectively represent the use of basic learner 1 at the t1th moment, t2th moment, t n Tracking prediction value at each moment; h(t1), h(t2), h(t n ) respectively represent the use of basic learner 2 at the t1th moment, t2th moment, and t n Tracking prediction value at a moment.

[0091] The third step is to use the three trained meta-learners g x (t), g y (t) and g θ (t) respectively for x t 、y t and θ t Perform tracking prediction to obtain the results of integrated learning, which are the three elements of the predicted state of the target: and

[0092] Step 7: Turn to detection.

[0093] Calculate the difference between the target's current and previous predicted azimuths

[0094]

[0095] in, and are the azimuth predictions for the current moment and the previous moment respectively.

[0096] Apply range adjustment to the difference:

[0097]

[0098] if The target is considered to be turned, Yes The prediction threshold is equal to the minimum steering angular velocity of the target multiplied by the sampling time interval of the sensor.

[0099] Step 8: Use Gaussian filter to predict the three elements of the target state and To perform smoothing:

[0100]

[0101] Where g(l) represents the value of the smoothed data at the lth position, 1≤l≤N, where N represents the total number of predicted states of the historical target, r(m) represents the value of the original data at the mth position, 1≤m≤N, and δ represents the standard deviation of the Gaussian filter, δ=2. is the normalization factor, ensuring that the total weight of the filter is 1.

[0102] Step 9: Calculate the target's current speed:

[0103]

[0104] in, Indicates the target's predicted speed at the current moment, Respectively represent the predicted position values ​​of the target on the x-axis at the current moment and the previous moment, They represent the predicted position values ​​of the target on the y-axis at the current moment and the previous moment respectively, and △t represents the sampling time interval, which is determined by the sampling frequency of the detector system.

[0105] Step 10: Get the predicted state of the target at the current moment And save the current prediction results.

[0106] Step 11, determine whether the current observation data is the last set of observation data, if so, execute step 12; otherwise, execute step 4.

[0107] Step 12: End tracking of the target.

[0108] The effects of the present invention will be further described below in conjunction with simulation experiments.

[0109] 1. Simulation experiment conditions.

[0110] The hardware test platform for the simulation experiment of the present invention is: the processor is CPU Intel Core i7-12700H, the main frequency is 2700MHz, and the memory is 16GB.

[0111] The software platform for the simulation experiment of the present invention is: Windows 10 official version, 64-bit operating system, PYTHON 3.8.

[0112] 2. Simulation content and results analysis:

[0113] The simulation experiment of the present invention sets up three scenarios, namely scenario 1, scenario 2 and scenario 3. The specific parameters of each scenario are shown in Table 1.

[0114] Table 1: Detailed information of parameter settings in this experiment

[0115]

[0116]

[0117] The simulation experiment of the present invention uses the present invention and two existing technologies (polynomial fitting least squares prediction method and extreme learning machine ELM regression prediction method) to fit and track the observation data of target motion for three different scenarios. The results are as follows: Figure 2 、 Figure 3 、 Figure 4 and Figure 5 The prediction results and performance results are compared.

[0118] In the simulation experiment, the two existing technologies used are:

[0119] The existing polynomial fitting least squares prediction method refers to the prediction method proposed by Gao Qiuying et al. in "Research on Least Squares Curve Fitting and Optimization Algorithm [J]. Industrial Control Computer, 2021, 34(11): 100-101.", referred to as the polynomial fitting least squares prediction method.

[0120] The existing technology extreme learning machine ELM regression prediction method refers to the regression prediction method proposed by Huang GB et al. in "Extreme learning machine for regression and multiclass classification[J].IEEE Transactions on Systems, Man, and Cybernetics, Part B(Cybernetics), 2011, 42(2): 513-529.", referred to as the extreme learning machine ELM regression prediction method.

[0121] The following combination Figure 2 、 Figure 3 、 Figure 4 and Figure 5 The effects of the present invention are further described.

[0122] The simulation results of scenario 1 are as follows: Figure 2 As shown:

[0123] Figure 2 (a) shows the result of using the present invention to predict the target movement, where the horizontal axis is the position of the target on the x-axis, and the vertical axis is the position of the target on the y-axis, and the units are both meters (m). In the simulation experiment of the present invention, an active detection domain is set. This area is a circle with a radius of 2000m. When the distance to the target is greater than 2000m, there will be no active detection, only passive detection, that is, only the azimuth information of the target is available. When the distance to the target is less than 2000m, both passive and active detection can be performed, that is, in addition to the azimuth information of the target, there is also the distance information of the target. From Figure 2 As can be seen from the enlarged figure in (a), the blue dotted line represents the actual motion trajectory of the target, the green dots represent the observed data, and the cyan dots represent the predicted results of the target motion by the present invention. The prediction of the present invention is closer to the target motion trajectory than the observed data, which shows that the present invention can effectively reduce the observation error. Note that there is no prediction result of the present invention near the starting point. This is because the present invention requires a certain solution cycle. Figure 2 (b) Figure 2 (c) and Figure 2 A similar situation occurs in (d).

[0124] Figure 2 (b) is the passive detection prediction result diagram of the present invention, which shows the change of the azimuth angle of the target during movement, where the horizontal axis is time, the unit is second (s), and the vertical axis is the direction angle of the target relative to the detection source, the unit is degree (°). Figure 2 In (b), the blue dotted line represents the actual azimuth angle change curve of the target, the green points represent the observed data, and the cyan points represent the prediction results of the target azimuth angle of the present invention. Figure 2 (b) It can be seen that after a certain solution period, the present invention begins to predict with high accuracy. Especially in the case of longer time periods, the prediction results still maintain a small error with the true value, demonstrating the stability and reliability of the present invention in long-term prediction.

[0125] Figure 2 (c) and Figure 2 (d) shows the target's motion trajectory in the x-axis direction and the motion trajectory in the y-axis direction, respectively. Their horizontal axes are time, and the unit is second (s). Figure 2 The vertical coordinate of (c) is the position of the target on the x-axis, and the unit is meter (m). Figure 2 The ordinate of (d) is the target's position on the y-axis, also in meters (m). It can be seen that the prediction results of the present invention in the x-axis direction and the y-axis direction are highly consistent with the true values, significantly better than the observed data. This further verifies the accuracy and reliability of the present invention in predicting target motion trajectories.

[0126] It can be seen that the intelligent tracking technology proposed in the present invention shows high accuracy and stability in the prediction of target motion trajectory by fusing active and passive information, and can effectively reduce observation errors.

[0127] The simulation results of scenario 2 are as follows: Figure 3 shown.

[0128] As shown in Table 1, the main difference between scenario 2 and scenario 1 is the speed of target movement, both the initial speed and the maximum speed are different, e.g. Figure 3 (a), where the horizontal axis is the target's position on the x-axis, and the vertical axis is the target's position on the y-axis, both in meters (m). Throughout the entire motion process, the target's motion speed is faster than in scenario 1, so the target has a longer trajectory. However, the prediction accuracy of the present invention is not affected, demonstrating relatively high accuracy.

[0129] Figure 3 (b) shows the change in the target's azimuth angle during motion, where the horizontal axis is time in seconds (s), and the vertical axis is the target's azimuth angle relative to the detection source in degrees (°). Throughout the entire motion process, the prediction results of the present invention are very close to the actual azimuth angle change of the target. Even when the target is moving at a high speed, the azimuth angle change of the target can still be accurately predicted. This shows that the present invention has high accuracy and stability in predicting the azimuth angle of high-speed moving targets.

[0130] Figure 3 (c) and Figure 3 (d) shows the target's motion trajectory in the x-axis and y-axis directions respectively. Their horizontal axes are time, and the unit is second (s). Figure 3 The vertical coordinate of (c) is the position of the target on the x-axis, and the unit is meter (m). Figure 3 The ordinate of (d) is the target's position on the y-axis, also in meters (m). It can be seen that even when the target is moving at a high speed, the prediction results of the present invention in the x- and y-axis directions are still highly consistent with the true values, significantly outperforming the observed data. This further demonstrates that the present invention can maintain high accuracy in predicting high-speed moving targets.

[0131] The simulation results of scenario 3 are as follows: Figure 4 shown.

[0132] As shown in Table 1, the main difference between scenario 3 and scenario 1 is the different initial distance of the target, e.g. Figure 4 (a) shows the target's position on the x-axis, and the ordinate is the target's position on the y-axis, both in meters (m). During the entire process of target motion, the target is outside the active detection domain, so the present invention only uses passive detection data for prediction.

[0133] Similarly, Figure 4 (c) and Figure 4 (d) There is no prediction of the target position by the present invention. Their horizontal coordinates are all time, and the unit is second (s). Figure 4 The vertical coordinate of (c) is the position of the target on the x-axis, and the unit is meter (m). Figure 4 The vertical coordinate of (d) is the position of the target on the y-axis, and the unit is also meter (m).

[0134] Figure 4 (b) shows the change of the target's azimuth angle during the movement, where the horizontal axis is time, the unit is seconds (s), and the vertical axis is the azimuth angle of the target relative to the detection source, the unit is degrees (°). Figure 4 As can be seen in (b), the blue dashed line represents the target's true azimuth curve, the green dots represent observed data, and the cyan dots represent the predictions made by the present invention. The cyan dots are closer to the blue dashed line than the green dots, indicating that at long distances, while the azimuth predicted by the present invention fluctuates slightly, its overall trend is relatively close to the true value.

[0135] In addition to predicting the position and azimuth of the target, the present invention can also track and predict the speed of the target and perform maneuver detection. Figure 5 The change of target speed over time and the maneuver detection result diagram are given.

[0136] Figure 5 (a) shows the target speed curve over time, where the horizontal axis is time in seconds (s), and the vertical axis is the target speed in meters per second (m / s). The blue dashed line represents the actual target speed change, and the cyan dots represent the prediction results of the present invention. It can be seen that the present invention can well track the target speed change trend, and the prediction results are consistent with the actual values ​​in most time periods, indicating that the speed prediction has high accuracy.

[0137] Figure 5(b) shows the target's turning information, where the horizontal axis is time in seconds (s), and the vertical axis is the target's turning information, that is, whether the target is turning at the current moment and the direction of the turning. The change in the turning information can reflect whether the target's movement direction has changed. Figure 5 As can be seen in (b), the target made frequent direction changes in most time periods, especially in the period from 40 seconds to 80 seconds, where the turning information showed multiple counterclockwise and clockwise direction changes, indicating that the target made frequent maneuvers at this time.

[0138] In summary, through Figure 2 、 Figure 3 、 Figure 4 and Figure 5 It can be seen that the present invention can not only accurately predict the position and azimuth information of the target, but also effectively perceive speed changes and simultaneously detect maneuvering information such as acceleration and steering, which is of great significance for target tracking.

[0139] The prediction results of the three simulation experiments were evaluated using two evaluation metrics: root mean square error (RMS) and error reduction percentage. The following formula was used to calculate the RMS error and error reduction percentage for each method and the observed data. The statistical results are shown in Table 2.

[0140] The root mean square error (RMSE) is used to quantify the difference between the predicted value and the true value. The calculation formula is as follows:

[0141]

[0142] Among them, y i represents the true value, represents the predicted value, and n represents the number of samples.

[0143] The error reduction percentage is used to evaluate the improvement of different methods relative to the observed data. The calculation formula is as follows:

[0144]

[0145] Among them, RMSE obs Represents the RMSE of the observed data, RMSE method Indicates the RMSE of a specific method.

[0146] Table 2 Performance comparison of various methods in three scenarios

[0147]

[0148] Analysis of the data in Table 2 shows that in all three scenarios, the present invention significantly outperforms other methods in both mean square error and error reduction. Specifically, in Scenario 1, the mean square error of the present invention is 15.050, and the error reduction reaches 58.028%. In Scenario 2, the mean square error is 42.682, and the error reduction reaches 46.861%. In Scenario 3, the mean square error is 35.954, and the error reduction reaches 54.949%. These results show that the present invention has high accuracy and stability in various scenarios.

Claims

1. A moving target tracking method combining active and passive information fusion and ensemble learning, characterized in that: Two basic learners for tracking prediction are constructed using the least squares method and the extreme learning machine. The linear fitting capability of the least squares method is combined with the nonlinear fitting capability of the extreme learning machine through the stacking method. Active and passive information fusion technology is used to supplement the active detection data. Outside the range of active detection, only passive detection is used. Within the range of active detection, passive detection is used to correct and supplement the outliers of active detection. The tracking method includes the following steps: Step 1: Fuse the active and passive detection data of the target at the current moment: Step 1.1: Determine whether the distance between the target to be tracked and the detection source is greater than or equal to a threshold. If so, use only passive detection data to predict the target's azimuth at the current moment. Otherwise, execute step 1.2; the threshold is equal to the maximum distance at which the active detector can effectively capture the target; In step 1.2, if the difference between the azimuth of the current active detection data and the azimuth of the previous active detection data exceeds the dynamic threshold at the current moment, the active detection data is determined to be abnormal. The azimuth of the passive detection data is used to perform a fusion correction on the azimuth of the active detection data through weighted fusion. Step 1.3: Using the fused and corrected azimuth detection values, calculate the target's observed position on the x-axis and y-axis through vector decomposition, generating the target's observation data at the current moment. This observation information includes the target's two-dimensional position, distance, and azimuth. Step 2: Use the polynomial fitting least squares method to fit the observed data of the target at the current moment to obtain the basic learner 1; Step 3: Use the extreme learning machine to fit the observation data of the target at the current moment to obtain the basic learner 2; Step 4: Use the stacking method in ensemble learning to combine the prediction results of base learner 1 and base learner 2 to obtain the prediction tracking result of the target at the current moment, that is, the three elements of the target prediction state and Step 5: Prediction of the azimuth angle element based on the target prediction state at the current moment The prediction of the target azimuth angle element at the previous moment Calculating the difference between the two azimuth angle predictions. If the difference is greater than a prediction threshold, it is determined that the target has turned at the current moment; the prediction threshold is equal to the target's minimum turning angular velocity multiplied by the sensor's sampling time interval; Step 6: Use Gaussian filter to predict the three elements of the target state and Perform smoothing; Step 7, calculate the target's movement speed at the current moment; Step 8: Determine whether the current observation data is the last set of observation data. If so, execute step 9; Otherwise, go to step 2; Step 9: End tracking of the target.

2. The moving target tracking method combining active and passive information fusion and integrated learning according to claim 1, characterized in that: The active detection and passive detection data of the target in step 1 refer to the following: the active detection data is the target azimuth and distance information detected by the active detection sensor, and the passive detection data is the target azimuth information detected by the passive detection sensor.

3. The moving target tracking method combining active and passive information fusion and integrated learning according to claim 1, characterized in that: The current dynamic threshold value described in step 1.2 is determined by the following formula: Among them, ε t represents the dynamic threshold at the tth moment, d t Indicates the distance between the target and the detection source at the tth moment, d max Indicates the maximum distance that the active detector can actively detect, △θ max Indicates the maximum detection error of the active detector, and its value is determined by the calibration data of the detector.

4. The moving target tracking method combining active and passive information fusion and integrated learning according to claim 1, characterized in that: The weighted fusion correction method described in step 1.2 is to use the azimuth of the passive detection data to perform fusion correction on the azimuth of the active detection data through the following formula: in, Indicates the target azimuth after fusion correction at the tth moment, represents the azimuth of active detection at the tth moment, represents the azimuth of passive detection at the tth moment, and α represents the weight of active detection data in the fusion correction, and its value range is 0≤α≤1.

5. The moving target tracking method combining active and passive information fusion and integrated learning according to claim 1, characterized in that: The observed position of the target on the x-axis and y-axis in step 1.3 is obtained by the following formula: in, Respectively represent the observation position of the target on the x-axis and y-axis at the t-th moment, represents the active observation distance of the target at the tth moment, Represents the target azimuth after fusion correction at the tth moment.

6. The moving target tracking method combining active and passive information fusion and integrated learning according to claim 5, characterized in that: The steps for fitting the observed data of the target at the current moment using the polynomial fitting least squares method described in step 2 are as follows: The first step is to construct the matrix A as follows: Where A represents an n×(m+1) matrix, n represents the total number of observation data determined by the requirements of solution time and solution accuracy, m represents the highest order of polynomial fitting, t1, t2, t n Respectively represent the t1th moment, t2th moment, t n At each moment, calculate the elements of A; The second step is to calculate the unknown coefficients a, b, and c according to the following formula: a=(A T A) -1 A T x b=(A T A) -1 A T y c=(A T A) -1 A T i Among them, the superscript T represents the transpose operation, and the superscript -1 represents the inverse operation. Respectively represent the t1th moment, t2th moment, t n The observed position of the target on the x-axis at each moment; Respectively represent the t1th moment, t2th moment, t n The observed position of the target on the y-axis at each moment; Respectively represent the t1th moment, t2th moment, t n The target azimuth after fusion correction at each moment; The third step is to get f x (t), f y (t), f θ (t) Three polynomial fitting functions form the basic learner 1: Among them, f x (t), f y (t), f θ (t) represents the three state elements of the normalized target at the tth moment and The polynomial fitting function, a k 、b k 、c k They represent the undetermined coefficients of the kth polynomial order in the polynomial fitting function, and their value range is k=0,1,2,...,m.

7. The moving target tracking method combining active and passive information fusion and integrated learning according to claim 6, characterized in that: The steps described in step 2 to use the extreme learning machine to fit the observed data of the target at the current moment are as follows: The first step is to determine the relevant parameters of the extreme learning machine network structure, including the dimension of the input layer, the number of hidden layer neurons, and the type of activation function; The second step is to randomly initialize the hidden layer weights and bias parameters of the extreme learning machine network; The third step is to calculate the element h of the hidden layer output matrix H according to the following formula. ij : Among them, h ij Represents the element in row i and column j of the hidden layer output matrix H, t p Indicates the tth p At each moment, 1≤p≤n, the value of i is equal to the value of p, c q The center of the qth neuron can be determined by random selection method, randomly selected from t p Select a moment as c q The value of σ q represents the width parameter of the qth neuron, σ q =σ, σ is the average distance between all input samples, The fourth step is to solve the hidden layer output weight matrix β according to the formula: β=(H T H+λI) -1 H T z Among them, λ represents the regularization term to prevent the matrix from being irreversible, Cond represents the condition number of the matrix; z represents the observed data of the target, respectively. and Substituting in, I is the identity matrix; The fifth step is to train the extreme learning machine to obtain and About the extreme learning machine fitting function h at the tth moment x (t), h y (t) and h θ (t), and use it as the basic learner 2 to participate in the ensemble learning. Their fitting function form is: Among them, β x,q , β y,q , β θ,q Respectively represent the output weight matrix β and and The corresponding weight; h q (t) represents the output of the qth neuron in the hidden layer.

8. The moving target tracking method combining active and passive information fusion and integrated learning according to claim 7, characterized in that: The steps of the stacking method in ensemble learning described in step 4 are as follows: In the first step, a linear regression model is selected as the meta-learner g(t). The coefficient w of the meta-learner is solved by minimizing the sum of squared errors. The form of the meta-learner is: g(t)=w0+w1f(t)+w2h(t), Where w = (w0, w1, w2) T represents the coefficient of the meta-learner, w=(X T X) -1 X T z, z represent the observation data of the target, respectively and Substitute, f(t), h(t) represent the prediction results of base learner 1 and base learner 2 at the tth moment, respectively. For the target state element and Train a meta-learner g separately x (t), g y (t) and g θ (t), that is, for the target state element Use f x (t) and h x (t) is the base learner of the meta-learner, for the target state element Use f y (t) and h y (t) is the base learner of the meta-learner, for the target state element Use f θ (t) and h θ (t) serves as the base learner of the meta-learner; The second step is to construct a stacked feature matrix X, which contains the prediction results of the base learner 1f(t) and the base learner 2h(t) and the constant term: Among them, f(t1), f(t2), f(t n ) respectively represent the use of basic learner 1 at the t1th moment, t2th moment, t n Tracking prediction value at each moment; h(t1), h(t2), h(t n ) respectively represent the use of basic learner 2 at the t1th moment, t2th moment, and t n Tracking prediction value at each moment; The third step is to use the three trained meta-learners g x (t), g y (t) and g θ (t) respectively for x t 、y t and θ t Perform tracking prediction to obtain the results of integrated learning, which are the three elements of the predicted state of the target: and 9. The moving target tracking method combining active and passive information fusion and integrated learning according to claim 1, characterized in that: The smoothing process described in step 6 is accomplished by: Where g(l) represents the value of the smoothed data at the lth position, 1≤l≤N, where N represents the total number of predicted states of the historical target, r(m) represents the value of the original data at the mth position, 1≤m≤N, and δ represents the standard deviation of the Gaussian filter, δ=2. is the normalization factor, ensuring that the total weight of the filter is 1.

10. The moving target tracking method combining active and passive information fusion and integrated learning according to claim 1, characterized in that: The target's current speed in step 7 is obtained by the following formula: in, Indicates the target's predicted speed at the current moment, Respectively represent the predicted position values ​​of the target on the x-axis at the current moment and the previous moment, They represent the predicted position values ​​of the target on the y-axis at the current moment and the previous moment respectively, and △t represents the sampling time interval, which is determined by the sampling frequency of the detector system.

Images