A false track identification method based on deep neural network

Abstract

The application discloses a false track identification method based on a deep neural network, and through the method, false tracks can be effectively inhibited in a dense clutter environment. Specifically, the method can be used in a track starting stage to identify false tracks of new-born tracks and replace traditional rule judgment, thereby effectively improving false information and missing information of target information. On the other hand, the method can be used in a track maintaining stage to identify the updating state of target tracks in real time, quickly find the correlation error between target tracks and radar measurements, timely correct the correct target measurements, and improve the stability of target tracks. Through the above two aspects, false tracks can be effectively removed in the clutter environment, and stable and reliable target information can be obtained.

Description

Technical Field

[0001] This invention relates to the field of radar detection technology, and in particular to a method for identifying false flight paths based on deep neural networks. Background Technology

[0002] With the increasing complexity of modern battlefield environments, radar sensors face challenges from electromagnetic interference, weather changes, and other factors, resulting in numerous false measurements. This makes it unavoidable to generate false tracks during target trajectory establishment. Furthermore, the increasing maturity of radar collaborative networking technology allows for the processing of multi-radar data, introducing even more useless false measurements and further increasing the probability of false tracks, severely interfering with the acquisition of actual target intelligence. Therefore, effectively suppressing false tracks and establishing true and reliable target tracks has become a hot research topic in the field.

[0003] Currently, the problem of false track suppression is usually addressed by using rigid rules for identification. However, actual radars have measurement errors and track a wide variety of targets, resulting in significant differences in target track states. Conventional false track suppression rules are difficult to match, often resulting in false and missed targets, making it impossible to establish complete and reliable target intelligence. Summary of the Invention

[0004] This invention provides a method for identifying false tracks based on deep neural networks, in order to solve the problem that existing technologies cannot accurately identify false tracks.

[0005] This invention provides a method for identifying false flight paths based on deep neural networks. The method includes: extracting features at the image and data levels from preprocessed historical target flight paths; specifically, modeling the historical target flight paths using images, representing the trajectory of the historical target flight paths using binary images, designing a feature extraction neural network using convolutional neural networks to extract features at the image level, fitting various states of the historical target flight paths using multi-order fitting theory, and using the rate of change of the fitted curve as the feature extraction result at the data level of the historical target flight paths; training the parameters of a preset classification neural network using the proposed image-level features and data-level features to construct a deep neural network, and optimizing the constructed deep neural network; classifying the target flight paths to be analyzed online based on the optimized deep neural network to identify and eliminate false flight paths, and finally outputting true target intelligence information.

[0006] Optionally, the step of extracting features at the image and data levels from the preprocessed historical target tracks includes: forming target track states from the acquired radar-measured historical target tracks through data association or Kalman filtering algorithms, recording the target track states, and completing classification and labeling.

[0007] The target trajectory sample is transformed into an image using a two-dimensional grid model to form a binary image of the target trajectory. The convolutional neural network receives the binary image information and performs convolution and pooling processing to form an image feature input signal, so as to realize the feature extraction of the target trajectory at the image level.

[0008] Second-order fitting processing is performed on different state information in the target track sample to obtain the second-order fitting function for each state. The rate of change of the curve of the second-order fitting function is used as the data feature input signal to realize the feature extraction of the target track at the data level.

[0009] Optionally, the step of forming a target track state from the acquired historical target tracks measured by radar through data association or a Kalman filtering algorithm includes:

[0010] The acquired historical target tracks from radar measurements are processed offline, including track initiation, radar measurement correlation, and track filtering, to form the target track and record its status (B, T, X, Y, Z, V). x V y V z A x A y A z ); where B is the target track number, T is the target track time, (X,Y,Z) represent the target's eastward, northward, and celestial positions, respectively, (V x V y V z (A) represents the target's eastward velocity, northward velocity, and skyward velocity, respectively. x A y A z ) represent the target's eastward acceleration, northward acceleration, and skyward acceleration, respectively.

[0011] Optionally, recording the target track status and completing the classification labeling includes:

[0012] The true value data is used to correlate all the generated target track status information to determine whether the target track is real or fake. Then the target track status is recorded and classified. The true value data includes cooperative target GPS data and / or civil aircraft ADS-B data.

[0013] Optionally, after recording the target track status and completing the classification and labeling, and before performing image conversion on the target track samples using a two-dimensional raster model, the method further includes:

[0014] After completing the data labeling of all target tracks, each target track is segmented using a sliding window method. The initial number of segments is K, where K is a natural number greater than or equal to 5. The segmented target track states are used as training samples according to the preset sliding window step. At the same time, the track classification results are combined for classification and storage to form positive and negative samples for deep neural network training.

[0015] Optionally, the step of using a two-dimensional raster model to perform image transformation on the target trajectory sample to form a binary image of the target trajectory, and then having a convolutional neural network receive the binary image information and perform convolution and pooling processing to form an image feature input signal, so as to realize feature extraction of the target trajectory at the image level, includes:

[0016] A two-dimensional grid model is established. The two-dimensional grid model has four quadrants. The length range of the x-axis and y-axis is [-5000m, 5000m]. The total size is 10000m×10000m. The size of a single grid cell is 500m×500m. The number of grid cells is 20×20.

[0017] Calculate the grid position of the track sample, taking the first position state (X0, Y0, Z0) in the track sample as the origin of the grid coordinates, and denote the grid position as (0, 0). and Calculate the state at other locations in the sample Grid position (C) i C j After calculating the positions of all grid cells, mark them according to the preset symbols.

[0018] Based on the CA model, extrapolation prediction is performed on each track state at a preset interval, the extrapolated grid position is calculated, and it is marked according to a preset symbol to highlight the image features of false tracks;

[0019] Complete the grid connection between the sample location and the predicted location in chronological order, and mark the grids that the connection passes through and the grids that have not been passed, to form a binary map of the target track.

[0020] Optionally, the step of training the parameters of a preset classification neural network using the proposed features to construct a deep neural network includes:

[0021] The features extracted from the image level and the features extracted from the data level are input into the binary classification neural network for training, thus completing the construction of the deep neural network and obtaining the deep neural network.

[0022] Optionally, features extracted at the image level and features extracted at the data level are input into a binary classification neural network for training to complete the construction of a deep neural network, including:

[0023] The binary map sample of the target trajectory is input into the feature extraction neural network to perform feature extraction and iterative training. The parameters of the feature extraction neural network are calculated and constructed. The feature extraction neural network contains a convolutional layer and a pooling layer. The convolutional layer consists of four 11×11 convolutional filters. The output of the convolutional layer is fed into the pooling layer after passing through the ReLU function. The pooling layer uses 2×2 submatrix average pooling.

[0024] By using the least squares method to fit the state information of each target trajectory sample in a second-order manner, we establish y = f(x) = ax 2 The second-order equation +bx+c, involved in the fitting sequence: eastward velocity fitting sequence (x i (T),y i (V x )), Northbound velocity fitting sequence (x i (T),y i (V y )), celestial velocity fitting sequence (x i (T),y i (V z )), Eastward acceleration fitting sequence (x i (T),y i (A x )), Northbound acceleration fitting sequence (x i (T),y i (A y )), celestial acceleration fitting sequence (x i (T),y i (A z ), i = 1, ..., K, where K is a natural number, and a, b, and c are all second-order fitting coefficients;

[0025] After completing the second-order least squares fitting, the parameters of the eastward velocity fitting equation were obtained. Northbound velocity fitting equation parameters Parameters of the fitting equation for celestial velocity ,c Vz ), parameters of the eastward acceleration fitting equation Northbound acceleration fitting equation parameters And the parameters of the fitting equation for axial acceleration

[0026] Differentiating the second-order fitting equations yields the following six sets of derivative equations, including the derivative equation for the eastward velocity. Northbound velocity derivative equation The derivative equation of the upward velocity equation Eastward acceleration derivative equation Northward acceleration derivative equation The derivative equation of the acceleration in the sky

[0027] Based on the derivative equation, substituting the values ​​into the target trajectory time series T = (t1, ..., t...) K ), calculate the data feature vector of the target trajectory Among them, the eastward velocity characteristic is The northbound velocity characteristics are The characteristics of celestial velocity are: Eastward acceleration characteristics are Northward acceleration characteristics are The characteristics of celestial acceleration are:

[0028] By differentiating the second-order fitting equation of the target state, the sum of the rates of change of the historical states of each target track is calculated to determine the real target and the false target.

[0029] The deep neural network is trained using the backpropagation algorithm, and the convolutional filters in the feature extraction neural network and the connection weights in the classification neural network are iteratively calculated to complete the construction of the deep neural network.

[0030] Optionally, the online classification of the target track to be analyzed based on the optimized deep neural network to identify and eliminate false tracks includes:

[0031] The target trajectory status is formed by using data association or Kalman filtering algorithm on the current target trajectory;

[0032] Record the target's trajectory status and complete the classification and labeling;

[0033] The target trajectory sample is transformed into an image using a two-dimensional grid model to form a binary image of the target trajectory. The convolutional neural network receives the binary image information and performs convolution and pooling processing to form an image feature input signal, thereby realizing feature extraction of the target trajectory at the image level.

[0034] Second-order fitting processing is performed on different state information in the target track sample to obtain the second-order fitting function for each state. The curve change rate of the second-order fitting function is used as the data feature input signal to realize the feature extraction of the target track at the data level.

[0035] The extracted features are input into the deep neural network for calculation to obtain the true and false attributes of the current target trajectory, and false trajectories are eliminated.

[0036] Optionally, the classification neural network adopts a deep fully connected neural network, including an input layer, four hidden layers and an output layer.

[0037] The beneficial effects of this invention are as follows:

[0038] The method described in this invention can effectively suppress false tracks in cluttered environments. Specifically, the method can be used in two ways: firstly, during the track initiation stage, to identify false tracks in newly formed tracks, replacing traditional rule-based judgments and effectively improving the detection of false and missed target intelligence. Secondly, during the track maintenance stage, it can be used to identify the updated status of target tracks in real time, quickly detect errors in the correlation between target tracks and radar measurements, correct them promptly, and improve the stability of target tracks. Through these two aspects, false tracks can be effectively eliminated in cluttered environments, obtaining stable and reliable target intelligence information.

[0039] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0040] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0041] Figure 1 This is a flowchart illustrating a false track identification and processing method based on a deep neural network provided in an embodiment of the present invention.

[0042] Figure 2 This is a flowchart illustrating a false track identification and processing method based on a deep neural network provided in an embodiment of the present invention.

[0043] Figure 3 This is a schematic diagram of binary image conversion provided in the first embodiment of the present invention;

[0044] Figure 4 This is a schematic diagram of the classification neural network structure provided in the first embodiment of the present invention. Detailed Implementation

[0045] This invention addresses the problem of inaccurate identification of false flight tracks in existing methods by proposing a false flight track identification method based on deep neural networks. First, feature extraction is performed on historical target flight tracks after radar data processing. One approach is to model the historical target flight tracks using a binary image to represent the trajectory, and then design a feature extraction neural network based on convolutional neural network principles for feature extraction. Another approach is to fit the various states of the target flight track using multi-order fitting theory, determining the rate of change of the fitted curve as the feature extraction result of the target flight track. Then, these two feature signals are passed to a classification neural network for parameter training, completing the construction of the deep neural network. Finally, the trained deep neural network is used online to classify target flight tracks as true or false, accurately identifying and eliminating false flight tracks, and ultimately outputting accurate target intelligence information. The invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and do not limit the scope of the invention.

[0046] This invention provides a method for identifying false tracks based on deep neural networks. See [link to relevant documentation]. Figure 1 The method includes:

[0047] S101. Perform image-level and data-level feature extraction on the preprocessed historical target tracks;

[0048] That is, in this embodiment of the invention, image modeling is performed on the historical target trajectory, the trajectory of the historical target trajectory is represented by a binary image, a feature extraction neural network is designed with the aid of a convolutional neural network to perform feature extraction at the image level, and the multi-order fitting theory is used to fit each state of the historical target trajectory, and the rate of change of the fitting curve is used as the feature extraction result at the data level of the historical target trajectory.

[0049] It should be noted that the preprocessing described in this embodiment of the invention specifically includes deduplication and format unification of the historical target tracks, that is, removing duplicate historical target tracks and removing data with incorrect formats, etc. Those skilled in the art can set these parameters arbitrarily according to actual needs. This invention will not describe them in detail. In general, this embodiment of the invention performs data deduplication and removes data with unqualified formats from the target tracks before feature extraction, so as to ensure the effectiveness of subsequent feature data extraction, thereby ensuring the accuracy of training deep neural networks and ultimately effectively ensuring the accuracy of identifying and eliminating false tracks.

[0050] The feature extraction at the image and data levels for the preprocessed historical target tracks described in this embodiment of the invention includes:

[0051] The acquired historical target tracks measured by radar are used to form target track status through data association or Kalman filtering algorithm, and the target track status is recorded and classified.

[0052] The target trajectory sample is transformed into an image using a two-dimensional grid model to form a binary image of the target trajectory. The convolutional neural network receives the binary image information and performs convolution and pooling processing to form an image feature input signal, so as to realize the feature extraction of the target trajectory at the image level.

[0053] Second-order fitting processing is performed on different state information in the target track sample to obtain the second-order fitting function for each state. The rate of change of the curve of the second-order fitting function is used as the data feature input signal to realize the feature extraction of the target track at the data level.

[0054] This invention trains a pre-defined classification neural network using extracted features.

[0055] S102. The proposed image-level features and data-level features are used to train the parameters of the preset classification neural network, a deep neural network is constructed, and the constructed deep neural network is optimized.

[0056] The classification neural network used in this embodiment of the invention is a deep fully connected neural network, which includes an input layer, four hidden layers and an output layer. Of course, in specific implementation, those skilled in the art can also set up other network models according to actual needs, and this invention does not make specific limitations in this regard.

[0057] In specific embodiments, after training and constructing the deep neural network, the present invention needs to continuously optimize the deep neural network through training and in actual use to make the deep neural network more accurate, so as to more accurately identify and eliminate false tracks.

[0058] For optimization, you can further train the deep neural network using the accurate labeled features, or continuously optimize the deep neural network using actual data during subsequent use, and so on.

[0059] S103. Based on the optimized deep neural network, the target tracks to be analyzed are classified as true and false online to identify and eliminate false tracks, and finally output true target intelligence information.

[0060] In a specific embodiment, this invention utilizes an optimized deep neural network to perform online classification of target tracks for true and false data. First, the target track needs to be processed using data association or a Kalman filter algorithm to form a target track state. This state is then recorded for classification. A two-dimensional raster model is used to transform the target track samples into a binary image. A convolutional neural network receives this binary image information and performs convolution and pooling processing to generate an image feature input signal, enabling feature extraction at the image level. Second-order fitting is performed on different state information in the target track samples to obtain second-order fitting functions for each state. The curve change rate of the second-order fitting function is then used as the data feature input signal to achieve feature extraction at the data level. Finally, the extracted features are input into the deep neural network for calculation to obtain the true or false attributes of the current target track and to eliminate false tracks.

[0061] In other words, the embodiments of the present invention need to perform data processing on the target tracks to be analyzed, extract features at the image level and data level, and then input them into the optimized deep neural network to achieve the classification of the target tracks as true or false.

[0062] In other words, this embodiment of the invention first extracts features from historical target tracks after radar data processing. First, it models the historical target tracks using image models, representing the target track trajectory with a binary image. Then, using convolutional neural networks, it designs a feature extraction neural network to extract features. Second, it uses multi-order fitting theory to fit each state of the target track, determining the rate of change of the fitted curve as the feature extraction result of the target track. These two types of feature signals are then passed to a classification neural network for parameter training, completing the construction of the deep neural network. Finally, the trained deep neural network is used online to classify target tracks as true or false, identifying and eliminating false tracks, and outputting accurate target intelligence information.

[0063] See Figure 2 The method described in this embodiment of the invention specifically includes:

[0064] The acquired radar measurement data is processed offline, and the target trajectory status is formed through algorithms such as data association and Kalman filtering.

[0065] Record the target's trajectory status, complete the classification and labeling, and form training samples;

[0066] The target trajectory samples are transformed into a binary image using a two-dimensional raster model. A convolutional neural network receives this binary image information, performs convolution and pooling processing, and generates an image feature input signal, enabling feature extraction of the target trajectory at the image level.

[0067] For different state information in the target trajectory samples, second-order fitting processing is performed to obtain the second-order fitting function for each state. The rate of change of the curve of this second-order fitting function is the data feature input signal, realizing feature extraction of the target trajectory at the data level;

[0068] The image and data feature signals are input into a binary classification neural network for training, thus completing the construction of the deep neural network and ending the offline processing.

[0069] Online processing performs offline data processing on the target tracks generated in real time, generates labeled training samples, then completes image and data feature extraction, and inputs the extracted feature signals into a deep neural network to perform classification calculations, obtain recognition results, and thus clarify the true or false attributes of the target track.

[0070] Based on the identification results, false tracks are eliminated and true tracks are retained, thereby solving the problem of numerous false tracks in dense clutter environments.

[0071] Practice has proven that this method can effectively suppress false tracks in cluttered environments. On one hand, it can be used in the initial stage of track detection to identify false tracks emerging from the ground, replacing traditional rule-based judgments and effectively improving the accuracy of target intelligence. On the other hand, it can be used in the track maintenance stage to identify the updated status of target tracks in real time, quickly detect errors in the correlation between target tracks and radar measurements, correct them promptly, and improve the stability of target tracks. Through these two aspects, false tracks can be effectively eliminated in cluttered environments, resulting in stable and reliable target intelligence information.

[0072] The following will combine Figures 2-4 The method described in the embodiments of the present invention will be explained and illustrated in detail through a specific example:

[0073] See Figure 2 The false track identification process based on deep neural networks described in this embodiment of the invention is divided into offline processing and online processing. Offline processing involves iteratively training the deep neural network by substituting labeled samples into the network to calculate network parameters. Online processing involves real-time classification of the target track by inputting the feature extraction signal of the target track into the deep neural network for classification calculation to obtain the target track identification result, which is used for the removal of false tracks.

[0074] Specifically, the offline processing in this embodiment of the invention includes:

[0075] Step 1: Offline processing of radar measurements;

[0076] Offline processing is performed on measurement data from single or multiple radars, including track initiation, radar measurement correlation, and track filtering, to form the target track and record the target track status (B, T, X, Y, Z, V). x V y V z A x A y A z Here, B is the target track number, T is the target track time, (X,Y,Z) represent the target's eastward, northward, and celestial positions, respectively, and (V x V y V z (A) represents the target's eastward velocity, northward velocity, and skyward velocity, respectively. x A y A z ) represent the target's eastward acceleration, northward acceleration, and skyward acceleration, respectively.

[0077] Step 2: Data tagging;

[0078] Data labeling uses ground truth data (such as cooperative target GPS data, civil aircraft ADS-B data, etc.) to associate all target track status information generated in step 1, and to clarify whether the target track is real or fake.

[0079] Specifically, in this embodiment of the invention, the ground truth data is first interpolated, and then the target track status information is associated with the ground truth data. A spherical gate is used for association, with an association distance threshold set to 500 meters. The nearest neighbor (NN) algorithm is employed, considering the target track with the smallest Euclidean distance as the true track. Target tracks that cannot be associated at all are considered false tracks.

[0080] Step 3: Target track storage;

[0081] After labeling all target tracks, each target track is segmented using a sliding window method, with an initial number of segments of K (K≥5) and a sliding window step of 1. The segmented target track states serve as training samples, and are also classified and stored in conjunction with the track classification results to form positive and negative samples for deep neural network training.

[0082] Step 4: Image conversion;

[0083] First, a two-dimensional grid model is established. The grid model has four quadrants, with the length range of the x-axis and y-axis both being [-5000m, 5000m]. The total size is 10000m × 10000m, the size of a single grid cell is 500m × 500m, and the number of grid cells is 20 × 20.

[0084] Then, the grid positions of the track samples are calculated. Taking the first position state (X0,Y0,Z0) in the track sample as the origin of the grid coordinates, and recording the grid position as (0,0), the remaining position states in the sample are calculated using formulas (1) and (2). Grid position (C) i C j After all grid positions have been calculated, mark them as "1".

[0085]

[0086]

[0087] Secondly, to highlight the image features of false tracks, extrapolation prediction is performed on each track state based on the CA model at a 1-second cycle, the extrapolated grid position is calculated, and the grid is marked as "1".

[0088] Finally, grid lines are connected between the sample locations and the predicted locations in chronological order. Grids traversed by the connection are marked as "1", and grids not traversed are marked as "0", forming a binary map of the target trajectory. See details below. Figure 3 As shown.

[0089] Step 5: Constructing the image feature extraction neural network;

[0090] Binary image samples of the target trajectory are input into a feature extraction neural network to perform feature extraction and iterative training, calculating and constructing the network's parameters. This feature extraction neural network contains one convolutional layer and one pooling layer. The convolutional layer consists of four 11×11 convolutional filters. The output of the convolutional layer is passed through a ReLU function before entering the pooling layer. The pooling layer uses 2×2 submatrix average pooling.

[0091] Step 6: State fitting;

[0092] By using the least squares method to fit the state information of each target trajectory sample in a second-order manner, we establish y = f(x) = ax 2 The second-order equation is +bx+c. The fitting sequences involved are as follows: Eastward velocity fitting sequence (x... i (T),y i (V x )), Northbound velocity fitting sequence (x i (T),y i (V y )), celestial velocity fitting sequence (x i (T),y i (V z )), Eastward acceleration fitting sequence (x i (T),y i (A x)), Northbound acceleration fitting sequence (x i (T),y i (A y )), celestial acceleration fitting sequence (x i (T),y i (A z )), i=1,…,K.

[0093] After completing the second-order least squares fitting, the parameters of the eastward velocity fitting equation were obtained. Northbound velocity fitting equation parameters Parameters of the fitting equation for celestial velocity Eastward acceleration fitting equation parameters Northbound acceleration fitting equation parameters And the parameters of the fitting equation for axial acceleration

[0094] Step 7: Data Feature Extraction;

[0095] Differentiating the second-order fitting equation from step 6 yields the following six sets of derivative equations: Eastward velocity derivative equations Northbound velocity derivative equation The derivative equation of the upward velocity equation Eastward acceleration derivative equation Northward acceleration derivative equation The derivative equation of the acceleration in the sky

[0096] Based on the derivative equation, substituting the values ​​into the target trajectory time series T = (t1, ..., t...) K ), calculate the data feature vector of the target trajectory Among them, the eastward velocity characteristic is The northbound velocity characteristics are The characteristics of celestial velocity are: Eastward acceleration characteristics are Northward acceleration characteristics are The characteristics of celestial acceleration are:

[0097] By differentiating the second-order fitting equation for the target state, the sum of the rates of change of the historical states of each target's trajectory is calculated, reflecting the characteristics of real and false targets. Typically, the velocity and acceleration of real targets in the air change relatively steadily, while the rate of change of the trajectory state of false targets is larger. Therefore, targets with smaller feature vector values ​​are real targets, and vice versa.

[0098] Step 8: Construct the classification neural network;

[0099] The classification neural network uses a deep fully connected neural network, consisting of one input layer, four hidden layers, and one output layer. The input layer has 106 nodes, which are the outputs from steps 5 and 7. Each hidden layer has 300 nodes and uses the ReLU activation function. The output layer has 2 nodes and uses the Softmax activation function. Figure 4 This is a schematic diagram of the classification neural network structure according to an embodiment of the present invention. For details, please refer to... Figure 4 To understand.

[0100] Step 9: Calculate the parameters of the deep neural network;

[0101] The deep neural network is trained by backpropagation algorithm, and the convolutional filters in the feature extraction neural network and the connection weights in the classification neural network are calculated iteratively to complete the construction of the deep neural network.

[0102] As can be seen, the deep neural network constructed in this invention is a convolutional neural network, and its network structure is shown in Table 1.

[0103] Table 1. Deep Neural Network Structure

[0104]

[0105] In this embodiment of the invention, online processing is mainly used to complete the real-time identification and elimination of false flight paths for practical engineering applications. Specifically, it includes the following five steps:

[0106] Step 1: Online processing of radar measurements;

[0107] Upon receiving measurement data from a single or multiple radars, the system first performs data association processing between measurements and tracks to find the target track closest to the radar measurements. Then, it uses radar measurements to perform state prediction, measurement prediction, and state update of the target track, completing the filtering and update of the target track state. Finally, it stores the updated target track state information for subsequent target track maintenance and false track identification. The stored target track state is the same as in step 1 of the offline processing, and the stored information is (B, T, X, Y, Z, V). x V y V z A x A y A z ).

[0108] Step 2: Store historical target tracks;

[0109] Historical track status information is stored for each target track, and the number of historical frames to be stored, K (K≥5), is determined, with the value of K consistent with offline processing. The stored status information includes: track time T, track eastward velocity V. xnorthbound speed V y Horizontal velocity V z Eastward acceleration A x northbound acceleration A y Horizontal acceleration A of the trajectory z .

[0110] Step 3: Feature extraction based on both image and data drivers;

[0111] Once the number of target track storage frames reaches a set threshold, image feature extraction and data feature extraction are performed based on the historical state of the target track. The image feature extraction method is detailed in step 4 of the offline processing, where an image transformation algorithm is used to model the historical track state, forming a binary image of the target track. The data feature extraction method is detailed in steps 6 and 7 of the offline processing, where a second-order fit is performed on the velocity and acceleration states in the target track, and the rate of change of the fitted function curve is characterized to obtain the data feature extraction results.

[0112] Step 4: False track identification and processing based on deep neural networks;

[0113] Image and data feature extraction values ​​are used as input layers to the constructed deep neural network for computation, yielding recognition probabilities. Based on these probabilities, it is determined whether the target track is a false track, thus completing the false track identification process.

[0114] Step 5: False track removal;

[0115] The source of identified false tracks is determined. If the false track was generated by the track initiation algorithm, it is deleted directly. Otherwise, if it was generated by the track maintenance algorithm, the latest false track state is deleted, and data association is performed again until the correct target track state is formed.

[0116] In general, the method described in this embodiment of the invention can be used in the initial stage of a flight path to identify false targets in newly formed paths, replacing traditional rule-based judgments to effectively improve the detection of false targets and missing targets. It can also be used in the path maintenance stage to identify the updated status of target paths in real time, quickly detect errors in the correlation between target paths and radar measurements, and promptly correct and identify the correct target measurements, thereby improving the stability of target paths.

[0117] Table 2 shows the comparison results of the recognition accuracy between the traditional algorithm and the algorithm of this invention.

[0118]

[0119] As shown in Table 2, the algorithm in this paper adds image features to the data features, and the accuracy of false track identification can reach 98%, which is better than the traditional algorithm based on data features. In practical use, it can achieve better recognition results and reduce the generation of false tracks.

[0120] Although preferred embodiments of the invention have been disclosed for illustrative purposes, those skilled in the art will recognize that various modifications, additions, and substitutions are possible, and therefore the scope of the invention should not be limited to the embodiments described above.

Claims

1. A method for identifying false tracks based on deep neural networks, characterized in that, include: Feature extraction is performed on the preprocessed historical target trajectory at both the image and data levels. Specifically, image modeling is performed on the historical target trajectory, and the trajectory is represented by a binary image. A feature extraction neural network is designed using a convolutional neural network to extract features at the image level. The multi-order fitting theory is used to fit each state of the historical target trajectory, and the rate of change of the fitted curve is used as the feature extraction result at the data level of the historical target trajectory. The proposed image-level features and data-level features are used to train the parameters of a pre-defined classification neural network to construct a deep neural network. The constructed deep neural network is then optimized. The classification neural network is a deep fully connected neural network, which includes one input layer, four hidden layers and one output layer. Based on the optimized deep neural network, the target tracks to be analyzed are classified as true and false online to identify and eliminate false tracks, and finally output true target intelligence information. The step of training parameters of a predefined classification neural network using the proposed features to construct a deep neural network includes: Features extracted from the image level and features extracted from the data level are input into a binary classification neural network for training, thus completing the construction of a deep neural network. This deep neural network specifically includes: The binary map sample of the target trajectory is input into the feature extraction neural network to perform feature extraction and iterative training. The parameters of the feature extraction neural network are calculated and constructed. The feature extraction neural network contains a convolutional layer and a pooling layer. The convolutional layer consists of four 11×11 convolutional filters. The output of the convolutional layer is fed into the pooling layer after passing through the ReLU function. The pooling layer uses 2×2 submatrix average pooling. The least squares method is used to fit the state information of each target trajectory sample in a second-order manner, and a system is established. Second-order equations, involved in the fitting sequence: eastward velocity fitting sequence Northbound velocity fitting sequence , celestial velocity fitting sequence Eastward acceleration fitting sequence Northbound acceleration fitting sequence Horizontal acceleration fitting sequence , K is a natural number. , and All are second-order fitting coefficients; After completing the second-order least squares fitting, the parameters of the eastward velocity fitting equation were obtained. Parameters of the northbound velocity fitting equation Parameters of the celestial velocity fitting equation Parameters of the fitting equation for eastward acceleration Parameters of the fitting equation for northward acceleration Parameters of the fitting equation for axial acceleration ; Differentiating the second-order fitting equations yields the following six sets of derivative equations, including the derivative equation for the eastward velocity. Northbound velocity derivative equation The derivative equation of the upward velocity Eastward acceleration derivative equation Northward acceleration derivative equation The derivative equation of axial acceleration ; Substituting the derivative equations into the target trajectory time series... Calculate the data feature vector of the target trajectory Among them, the eastward velocity characteristic is The northbound velocity characteristics are The upward velocity characteristics are The eastward acceleration characteristic is The northward acceleration characteristic is The characteristics of celestial acceleration are ; By differentiating the second-order fitting equation of the target state, the sum of the rates of change of the historical states of each target track is calculated to determine the real target and the false target. The deep neural network is trained using the backpropagation algorithm, and the convolutional filters in the feature extraction neural network and the connection weights in the classification neural network are iteratively calculated to complete the construction of the deep neural network.

2. The method according to claim 1, characterized in that, The process of extracting features from the preprocessed historical target tracks at both the image and data levels includes: The acquired historical target tracks measured by radar are used to form target track status through data association or Kalman filtering algorithm, and the target track status is recorded and classified. The target trajectory sample is transformed into an image using a two-dimensional grid model to form a binary image of the target trajectory. The convolutional neural network receives the binary image information and performs convolution and pooling processing to form an image feature input signal, so as to realize the feature extraction of the target trajectory at the image level. Second-order fitting processing is performed on different state information in the target track sample to obtain the second-order fitting function for each state. The rate of change of the curve of the second-order fitting function is used as the data feature input signal to realize the feature extraction of the target track at the data level.

3. The method according to claim 2, characterized in that, The acquisition of historical target tracks measured by radar is used to form target track status through data association or Kalman filtering algorithms, including: The acquired historical target tracks from radar measurements are processed offline, including track initiation, radar measurement correlation, and track filtering, to form the target track and record its status. ;in, For the target track number, For the target trajectory time, These represent the target's eastward, northward, and celestial positions, respectively. These represent the target's eastward velocity, northward velocity, and skyward velocity, respectively. These represent the target's eastward acceleration, northward acceleration, and upward acceleration, respectively.

4. The method according to claim 2, characterized in that, The process of recording the target trajectory status and completing classification and labeling includes: The true value data is used to correlate all the generated target track status information to determine whether the target track is real or fake. Then the target track status is recorded and classified. The true value data includes cooperative target GPS data and / or civil aircraft ADS-B data.

5. The method according to claim 2, characterized in that, After recording the target track status and completing the classification and labeling, and before performing image conversion on the target track samples using a two-dimensional raster model, the method further includes: After completing the data labeling of all target tracks, each target track is segmented using a sliding window method, with an initial number of segments. According to the preset sliding window step, the target trajectory state is segmented as training samples, and the trajectory classification results are combined for classification and storage to form positive and negative samples for deep neural network training.

6. The method according to claim 5, characterized in that, The process involves using a two-dimensional raster model to transform the target trajectory sample into an image, forming a binary image of the target trajectory. A convolutional neural network receives this binary image information, performs convolution and pooling processing, and then generates an image feature input signal to achieve feature extraction of the target trajectory at the image level. This includes: A two-dimensional raster model is established, which has four quadrants. shaft and The length range of the shaft is all Total size is The size of a single grid cell is The number of grid cells is ; Calculate the grid position of the track sample, starting with the first position state in the track sample. Let be the origin of the raster coordinate system, and denote the raster position as . ,pass and Calculate the state at other locations in the sample grid position After calculating the positions of all grid cells, mark them according to preset symbols; Based on the CA model, extrapolation prediction is performed on each track state at a preset interval, the extrapolated grid position is calculated, and it is marked according to a preset symbol to highlight the image features of false tracks; Complete the grid connection between the sample location and the predicted location in chronological order, and mark the grids that the connection passes through and the grids that have not been passed, to form a binary map of the target track.

7. The method according to any one of claims 1-6, characterized in that, The online classification of target tracks for analysis based on the optimized deep neural network to identify and eliminate false tracks includes: The target trajectory status is formed by using data association or Kalman filtering algorithm; Record the target's trajectory status and complete the classification and labeling; The target trajectory sample is transformed into an image using a two-dimensional grid model to form a binary image of the target trajectory. The convolutional neural network receives the binary image information and performs convolution and pooling processing to form an image feature input signal, thereby realizing feature extraction of the target trajectory at the image level. Second-order fitting processing is performed on different state information in the target track sample to obtain the second-order fitting function for each state. The curve change rate of the second-order fitting function is used as the data feature input signal to realize the feature extraction of the target track at the data level. The extracted features are input into the deep neural network for calculation to obtain the true and false attributes of the current target trajectory, and false trajectories are eliminated.

8. The method according to any one of claims 1-6, characterized in that, The method further includes: preprocessing the historical target trajectory; The preprocessing of the historical target track includes: The historical target tracks are deduplicated and formatted.

Images