A method and device for error registration of a multi-radar system based on LSTM

Abstract

The application discloses a kind of based on LSTM's multi-radar system error registration method and device, obtain target track, target track is located in the overlapping region of multiple radar detection regions in multi-radar system;According to target track, generate the feature sequence of each time;With the feature sequence of each time as input data, determine the position information of target in the overlapping region of each time by detection region classification network;Determine the system error of each radar in multi-radar system based on position information;The present application can greatly improve the precision of system error by assigning different system errors to the overlapping region of multiple radar detection regions according to different word regions;At the same time, through the detection region classification network, the position information of target track can be corrected to obtain more accurate target position information, thereby further improving the estimation precision of system error.

Description

Technical Field

[0001] This invention belongs to the field of information fusion technology, and in particular relates to a method and apparatus for error registration of a multi-radar system based on LSTM. Background Technology

[0002] Multi-radar target tracking utilizes target information acquired from multiple sensors, processing it through algorithms at different levels and fusion stages to obtain a more accurate estimate than that obtained from a single sensor. Since its inception, this concept has been widely applied in both military and civilian fields. Military applications include missile interception, aircraft reconnaissance and tracking, early warning and penetration, and real-time battlefield monitoring. Civilian applications encompass ground-based human tracking, traffic management, air traffic control, maritime vessel monitoring, and currently popular applications such as unmanned vehicle driving, facial recognition, and gesture tracking in computer vision. Radar, with its all-weather operation, has become a crucial sensor in strategic defense systems, and target tracking based on system error registration has become a highly sought-after technology.

[0003] Systematic error registration is the first step in track fusion tracking. Systematic error is a deterministic error present in radar observation systems that cannot be removed by filtering. It requires prior estimation and compensation, and is the primary prerequisite and key task for track fusion target tracking. If the track cannot correctly compensate for systematic errors, not only will the target tracking error be larger than the theoretical value and the tracking effect of multiple radars be worse than that of a single radar, but it may also lead to multiple tracks for the same target, negating the advantages of the radar network system itself and significantly impacting track tracking and subsequent track fusion and situation assessment tasks.

[0004] Currently, most traditional multi-radar systems rely on a constant system error, which leads to a significant decrease in the accuracy of error estimation. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for error registration of multi-radar systems based on LSTM, so as to improve the estimation accuracy of system errors.

[0006] This invention adopts the following technical solution: an error registration method for a multi-radar system based on LSTM, comprising the following steps:

[0007] Acquire the target track, which is located in the overlapping area of ​​multiple radar detection zones in a multi-radar system;

[0008] Generate a feature sequence for each time step based on the target trajectory;

[0009] Using the feature sequence at each time step as input data, the location information of the target in the overlapping area at each time step is determined by a probe region classification network.

[0010] The system error of each radar in a multi-radar system is determined based on location information.

[0011] Preferably, generating the feature sequence for each time step based on the target trajectory includes:

[0012] Generate the first target position coordinates at each moment in the multi-radar system coordinate system based on the target trajectory;

[0013] Generate the second target position coordinates at each moment in the polar coordinate system based on the position coordinates of each radar and the target position coordinates in the multi-radar system;

[0014] By combining the coordinates of the second target position corresponding to each radar at each time step, the feature sequence of each time step is obtained.

[0015] Preferably, the detection region classification network is an LSTM network, specifically including an LSTM layer, a fully connected layer, a softmax layer and a classification layer connected in sequence;

[0016] The input dimension of the LSTM layer is the same as the dimension of the feature sequence.

[0017] Preferably, the system error of each radar in a multi-radar system is determined based on location information:

[0018] Determine the sub-region number of the target within the overlapping area based on the location information;

[0019] The system error table for each radar is queried based on the sub-region number, and the system error corresponding to the sub-region number in the system error table for each radar is determined.

[0020] Preferably, the system error table is generated based on the radar's radial distance and azimuth angle.

[0021] The preferred method for training an LSTM network is as follows:

[0022] Generate realistic target tracks;

[0023] Different system errors are added to the real target track based on the system error table of each radar to obtain the detection track of each radar;

[0024] Generate corresponding polar coordinate values ​​based on the detection tracks of each radar;

[0025] By combining the polar coordinate values ​​corresponding to each radar at each time step, the training feature sequence at each time step is obtained;

[0026] Combine the training feature sequences at each time step to generate the training sample set for the LSTM network.

[0027] Preferably, the loss function of the LSTM network is the cross-entropy classification loss function.

[0028] Another technical solution of the present invention: an LSTM-based multi-radar system error registration device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned LSTM-based multi-radar system error registration method.

[0029] The beneficial effects of this invention are: by assigning different system errors to the overlapping areas of multiple radar detection areas according to different word regions, this invention can greatly improve the accuracy of system errors; at the same time, the position information of the target trajectory can be corrected through the detection area classification network to obtain more accurate target position information, thereby further improving the accuracy of system error estimation. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the error registration method for a multi-radar system based on LSTM according to the present invention.

[0031] Figure 2 This is a schematic diagram of the architecture of the detection region classification network according to an embodiment of the present invention;

[0032] Figure 3 This is a schematic diagram of the system error values ​​within the sector-shaped area blocks of different detection ranges for different radars in the simulation verification embodiment of the present invention;

[0033] Figure 4 This is a schematic diagram of sub-region division in a simulation verification embodiment of the present invention;

[0034] Figure 5 The graph shows the RMSE results of the system error in the simulation verification embodiment of the present invention. Detailed Implementation

[0035] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0036] With the continuous development of machine learning and deep learning, data-driven methods for system error registration have emerged in recent years. Due to the limited information content and temporal nature of radar point data, these methods often cannot exist independently. The general approach is to combine a network with corresponding filters, using the network to input the constructed feature data into a traditional machine learning or deep learning network to estimate the system error. This approach is currently still in the exploratory stage, and no universally accepted and mature method has yet emerged.

[0037] Therefore, in order to address some shortcomings of the above methods, this invention designs a multi-radar system error registration method based on LSTM. This method leverages the advantage of deep learning in simulating complex nonlinear mappings and comprehensively utilizes the information interaction of multi-radar track data to improve the accuracy of system error classification and discrimination, thereby reducing the inaccuracy of system error estimation when considering spatial distribution structures.

[0038] System error registration is the primary task of radar point fusion tracking. Traditional methods often require sufficient prior information and accurate prior parameters. System error registration methods with known and unknown prior models often lead to problems such as unstable system error values ​​and system divergence.

[0039] Therefore, this invention considers using a data-driven approach, combined with an LSTM deep learning network, to complete the system error registration task. The research background focuses on system bias registration under the condition that the sensor system error is a dynamically changing unknown quantity that evolves over time. The sensor system bias is divided into sector-shaped spatial regions and labeled with different bias values ​​to simulate the dynamically changing unknown quantity over time. This deep learning network, by comprehensively utilizing the information interaction of multi-sensor data, is expected to significantly improve the accuracy of system bias discrimination and complete the system error registration task.

[0040] This invention discloses a multi-radar system error registration method based on LSTM, comprising the following steps: acquiring a target track, wherein the target track is located in the overlapping area of ​​multiple radar detection areas in the multi-radar system; generating a feature sequence for each time step based on the target track; using the feature sequence for each time step as input data, determining the target's position information in the overlapping area at each time step through a detection area classification network; and determining the system error of each radar in the multi-radar system based on the position information.

[0041] This invention significantly improves the accuracy of system errors by assigning different system errors to the overlapping areas of multiple radar detection zones according to different sub-regions. At the same time, the detection zone classification network can correct the target trajectory position information to obtain more accurate target position information, thereby further improving the accuracy of system error estimation.

[0042] The detection region classification network in this embodiment of the invention is an LSTM network, such as... Figure 2 As shown, it specifically includes an LSTM layer, a fully connected layer, a softmax layer, and a classification layer connected in sequence; the input dimension of the LSTM layer is the same as the dimension of the feature sequence.

[0043] First, the LSTM network needs to be trained. The specific training method is as follows: generate real target tracks; add different system errors to the real target tracks based on the system error table of each radar to obtain the detection tracks of each radar; generate corresponding polar coordinate values ​​based on the detection tracks of each radar; combine the polar coordinate values ​​of each radar at each time step to obtain the training feature sequence at each time step; combine the training feature sequences at each time step to generate the training sample set of the LSTM network. During the training process, the loss function of the LSTM network is the cross-entropy classification loss function.

[0044] More specifically, LSTM network training methods include... Figure 1 As shown, this is achieved through the following steps.

[0045] First, a portion of the generated real target tracks is used for training, and the other portion is used for testing. Then, measurement noise and system errors are added to the real target tracks. Considering that system errors are related to the spatial distribution of radar areas, this embodiment uses a numerical assignment method instead of a radar-detection-based system error model to construct a spatial distribution model of system errors, i.e., a system error table.

[0046] Specifically, the system error table is generated based on the radar's radial distance and azimuth angle. The radar detection area is divided into different sub-regions by a grid within a sector based on different radial distances and azimuth angles. Different radial distance and azimuth angle system errors are assigned to each sub-region. When traversing different sector grid regions, the local track data has the system error of the corresponding sector grid region (i.e., sub-region) added to it. At different times, as the radial distance and azimuth angle between the track and the radar change, the value of the system error also changes accordingly.

[0047] Furthermore, a feature sequence for each moment is generated based on the track with added systematic errors. Specifically, the first target position coordinates at each moment are generated in the multi-radar system coordinate system based on the target track; the second target position coordinates at each moment are generated in the polar coordinate system based on the position coordinates of each radar in the multi-radar system and the target position coordinates; the second target position coordinates corresponding to each radar at each moment are combined to obtain the feature sequence for each moment. The above feature sequence generation process is also applied to the specific systematic error registration process.

[0048] After obtaining the feature sequence at each time step, the feature sequences of all time steps of the track are combined and input into the detection area classification network, which is the LSTM network.

[0049] In this embodiment, for ease of understanding, the multi-radar system is defined as two radar systems. The LSTM network takes a sequence of size 4 (the number of features in the input data) as input, specifies an LSTM layer containing 128 hidden units, and outputs the complete sequence. Finally, the network includes a fully connected layer of size 19, followed by a softmax layer and a classification layer to specify 19 classes (corresponding to 19 sub-regions). The processed feature sequence passes through a single LSTM layer, then through a fully connected layer to extract features of the spatial distribution model, then through a SOFTMAX layer to express the feature results in probabilistic form, and finally through a classification layer to classify the probabilistic results, thus completing the training of the network.

[0050] In the field of deep learning, LSTM is a classic and mature recurrent neural network, often used to classify sequential data and learn the long-term dependencies between time steps of sequential data. As more and more machine learning methods with deep learning at the forefront are being developed, they are gradually being applied more and more in the research of time series analysis data classification, and have gained widespread recognition and application in the industry.

[0051] The main purpose of this invention using LSTM is to classify multi-radar track sequence data at each moment. The track data of two radars for a target are simultaneously input into the network. The data first passes through an LSTM layer containing multiple hidden layers, outputting complete track sequence data. Then, it passes through a fully connected layer, a SOFTMAX layer, and a classification layer to extract the sub-region where the target is located at each moment (i.e., pattern classification).

[0052] LSTM, as a special type of RNN network, introduces CELL cells, which have three structures that can adjust information: forget gate structure, input gate structure, and output gate structure.

[0053] Long Short-Term Memory (LSTM) networks introduce memory units and various gating units to address the long-term information retention problem inherent in classic recurrent neural network models. LSTM controls the state of the memory units through input gates, forget gates, and output gates. These three gates are all fully connected layers with sigmoid activation functions. The current input and the hidden state from the previous time step are respectively entered into these three gates, resulting in three outputs in the range (0,1), which are used to selectively add or delete information from the memory units. The specific computation process is as follows:

[0054] I t =σ(X) t W xi +H t-1 W hi + bi ),

[0055] F t =σ(X)t W xf +H t-1 W hf +b f ),

[0056] O t =σ(X) t W xo +H t-1 W ho +b o) ,

[0057] Among them, X t It is the input, H t-1 It is the hidden state from the previous moment, I t F t O t These are the input gate, forget gate, and output gate, respectively, and σ is the sigmoid activation function. (Using C...) t Representing memory units, using This represents the information to be added to the memory unit at the current moment, calculated using the following formula.

[0058]

[0059]

[0060] The formula shows that the state of a memory cell at time t consists of two parts: one part comes from the memory cell at the previous time step, and the forget gate determines how much historical information to retain; the other part comes from the calculated current state, and the input gate determines how much current information to add to the memory cell. By using addition operations instead of direct substitution operations, Long Short-Term Memory (LSTM) networks can better capture long-range dependencies in sequences.

[0061] These gate structures can learn which information in sequence data is valid and needs to be retained, and which is unnecessary and needs to be deleted. The advantage of this is that it can convey relevant information from long sequence data to make predictions.

[0062] The LSTM network extracts features from the spatial distribution structure of system errors of multiple radars (i.e., the overlapping detection area composed of sub-regions) and fuses the spatial distribution features of multiple radars (i.e., the features of each sub-region) to enhance the feature extraction of spatial overlap of multiple radars.

[0063] A sequence-to-sequence LSTM deep neural network is used, containing a fully connected layer, followed by a SOFTMAX layer and a classification layer. The fully connected layer maps the learned "distributed feature representation" to the sample label space. The SOFTMAX layer performs a transformation on the deep learning network's output, representing it as a probability. The classification layer categorizes the probabilities. Finally, location data from multiple radars detecting the same target are input together, allowing for information exchange among the radar data. This trains the deep neural network to classify the sequence data at each time step, outputting the sub-region where the target is located at each time step from the multiple radar sequences. Thus, the network completes the task of registering systematic errors under spatial distribution.

[0064] To train the aforementioned deep neural network, a large number of training samples need to be constructed. When constructing the training samples, the target moves from any position, along any direction, and within a certain speed range within the detection range of two radar fields of view. The purpose is to simulate radar point measurement data. The generated radar point data (x1, y1), (x2, y2) in the Cartesian coordinate system are converted into radial distance and azimuth angle (ρ1, θ1), (ρ2, θ2) in the polar coordinate system, and the radial distance system error and azimuth angle system error values ​​within their corresponding spatial range are added. The specific formula is as follows, where X... k It is the state vector of the system model.

[0065]

[0066]

[0067] Among them, X k (1) Represents the x-coordinate of the target in a rectangular coordinate system, X k (3) Represents the vertical coordinate of the target in a rectangular coordinate system.

[0068] like Figure 3 As shown, in this embodiment, the detection ranges of the two radars are divided into five annular blocks according to the radial distance. Each annular block is further divided into three regions according to different azimuth angles. The detection ranges of the two radars largely overlap, and the regions within the overlapping range also overlap. Therefore, the target data at each moment is classified into 19 categories (i.e., there are 19 overlapping regions). The network needs to complete the multi-classification task of the targets, and the loss function of the entire network is composed of the cross-entropy classification loss function.

[0069] For the loss function, we choose the cross-entropy loss function to characterize the deviation between the network's classification result and the true value. The standard form of the cross-entropy loss function is as follows:

[0070]

[0071] x, y, n represents the training sample index, the true label of the training sample, the predicted label of the LSTM network, and the total number of training samples, respectively. The sample labels should be represented in an encoded format, and the predicted labels of the LSTM network should be the probability distributions of different classes.

[0072] Next, the two radar training samples, processed by the system error module, are input together into the deep neural network for information interaction. The trained LSTM deep neural network will distinguish the boundaries of the fan-shaped regions within the multi-radar detection range. Considering the errors introduced by the network itself, it may not be able to estimate the system error of measurement data outside the boundaries of the multi-radar fan-shaped detection area. Therefore, this setting eliminates tracks outside the radar detection range boundaries to avoid affecting the network training.

[0073] More specifically, a total of 5,000 local tracks were generated in the training samples. After data preprocessing and processing by the system error module, the two radars generated a total of 10,000 local track data. The feature sequence corresponding to each time step was input into the LSTM deep neural network and the complete sequence was output. Each input sequence has four features, which correspond to the radial range and azimuth angle of the two radars detecting the same target. The four features are the radial range and azimuth angle of radar A and the radial range and azimuth angle of radar B.

[0074] A uniform linear motion (CV) model is used as the dynamic evolution model during track generation. The target's motion equations and measurement equations are as follows:

[0075] X(k+1)=F·X(k)+μ(k),

[0076]

[0077] Where X(k+1) represents the target's state value at time (k+1), F is the state transition matrix, X(k) represents the target's state value at time k, μ(k) represents the system noise, Z(k) represents the sensor's measurement value at time k, H is the measurement matrix, and B(k) represents the sensor's system error value at time k. Represents measurement noise. x represents the target's coordinates in the X direction. y represents the target's velocity in the X direction, and y represents the target's coordinates in the Y direction. This represents the target's velocity in the Y direction. The X and Y directions are the x-axis and y-axis directions in the Cartesian coordinate system where the multi-radar system is located.

[0078] The state transition matrix and measurement matrix are as follows:

[0079]

[0080]

[0081] Where T represents the radar's scanning period.

[0082] Finally, the multi-radar point data is classified to obtain the sub-region number of the target at each time moment. Then, based on the location of the sub-region, the specific system error value is obtained by looking up the system error table. The radar point data is then registered to eliminate errors caused by the sensor itself, and a filter is applied to minimize the inaccuracy of local track target tracking due to measurement noise. The training process is now complete.

[0083] In practical applications, the sub-region number of the target in the overlapping area is determined based on the location information; the system error table of each radar is queried based on the sub-region number, and the system error corresponding to the sub-region number in the system error table of each radar is determined.

[0084] In addition, to illustrate the effectiveness of the method of the present invention, the following simulation verification examples were performed:

[0085] A 2D multi-radar spatial distribution trajectory simulation environment is set up. Two radars have a detection range of 0-100000m in a Cartesian coordinate system, and their radar positions are (0,0) and (10,0) respectively. 5 The scan period is T = 1s. Assuming the target moves at a constant velocity in a straight line during the trajectory phase, with a speed range of 30m / s to 50m / s, a target is set up, and its initial position coordinates are randomly generated within the two-dimensional detection range plane. The detection areas of the two radars are as follows: Figure 3 As shown, the system error of a single radar in each region is fixed. By setting different system errors, the corresponding system error values ​​for each sub-region within the radar's detection range can be obtained.

[0086] Multiple Monte Carlo simulations were conducted on the above tracks (each Monte Carlo simulation generates a new track based on the corresponding error value). The root mean square error (RMSE) of the estimated radial distance and azimuth position was used as an indicator to evaluate the performance of the tracking method. The calculation formula is shown below, where... Let r be the radial distance estimate and azimuth estimate after the i-th Monte Carlo simulation at time k in a certain dimension. i,k θ i,kHere, represents the true values ​​of radial distance and azimuth angle, and MC represents the number of Monte Carlo simulations. From this, we can obtain the RMSE of each track at each time step. The lower the value, the more advantageous the tracking method is.

[0087]

[0088]

[0089] The simulation scenario sets the measurement noise and system error values ​​in a two-dimensional plane, with the standard deviation of radial distance measurement noise being 100m and the standard deviation of azimuth measurement noise being 0.003rad.

[0090] Table 1 shows the radial distance system error values ​​in the simulation scenario, and Table 2 shows the azimuth angle system error values ​​in the simulation scenario.

[0091] Table 1

[0092]

[0093] Table 2

[0094]

[0095] This method uses a numerical assignment method instead of an error model based on radar detection principles. It divides the radar detection area into grids and sets different radial distance deviations and azimuth deviations on the grids, which reduces the difficulty of establishing a complex radar system error model. The regional distribution of system errors is obtained by using the assignment method.

[0096] The radar detection area is divided into grids, and a region D is designated to cover the area where the system error to be estimated is located, such as... Figure 4 As shown. The range of radar space detection is [d0, d... m ],0<d0<d m Where, distance refers to the distance to the radar, d0 is the minimum distance, d m This represents the maximum distance. Its azimuth range is [β0, β...]. n ], where 0 < β0 < β n <360 represents the angle with true north; take d0 < d1 ... < d m Divide the distance range into m parts (m = 5 in this embodiment), and take β0 < β1 ... < β n The azimuth range is divided into n (n=3 in this embodiment), and the vertices of the grid are called nodes, denoted as d from the radar distance. i Taking radar B as an example, if the radial distance of the radar point is between d0 and d1, then the radial distance system error of the radar point is assigned a value of 150m. If the azimuth angle is between β0 and β1, then the azimuth angle system error of the radar point is assigned a value of 0.229°.

[0097] RMSE results Figure 5 show, Figure 5 (a) and Figure 5 (b) These figures represent the radial distance RMSE and azimuth RMSE results of radar B estimated by the LSTM network after 1000 Monte Carlo simulations of a single track (sampling period of 1s, totaling 1000 points). The horizontal axis represents 1000 sampling points, and the vertical axis represents the RMSE value (a smaller RMSE indicates better performance). According to these two figures, the maximum azimuth deviation RMSE of radar B is only 0.000991464, and the maximum radial distance deviation RMSE is only 9.91464, demonstrating very high accuracy.

[0098] Figure 5 Figures (c) and (d) represent the radial distance RMSE and azimuth RMSE results of radar A, estimated by the network after 1000 Monte Carlo simulations of a single track (sampling period of 1 second, with a total of 1000 sampling points). The horizontal axis represents 1000 sampling points, and the vertical axis represents the RMSE value (a smaller RMSE indicates better performance). According to these two figures, the maximum azimuth deviation RMSE of radar A is only 0.000752994, and the maximum radial distance deviation RMSE is only 7.52994, demonstrating higher accuracy than radar A.

[0099] Therefore, the LSTM-based multi-radar system error registration method demonstrates high accuracy in system error estimation and can effectively estimate and register multiple system errors in spatially distributed scenarios. In summary, this invention proposes an LSTM-based system error registration method that constructs a spatial distribution model of system errors by dividing them into sector-shaped spatial regions. Utilizing data-driven deep learning, it comprehensively considers the spatial distribution information of the target, leveraging the advantages of deep learning to significantly reduce the error tolerance of system error estimation. Compared to traditional system error registration methods, which only require coarse adjustment of hyperparameters based on the magnitude of system error and measurement noise, this method simplifies the selection and tuning process of hyperparameters in traditional methods. It also uses a large amount of radar data to replace prior information, overcoming the difficulty in obtaining precise prior parameters.

[0100] This invention constructs a spatial distribution model of sensor system errors by dividing the system into sector-shaped spatial regions. Different regions are marked with different error values ​​to simulate the dynamic and unknown changes of system errors over time. By comprehensively utilizing the interaction of data from multiple sensors, the traditional method's requirement for precise prior parameters is reduced. The spatial distribution information of the target is comprehensively considered, and the error tolerance of system error estimation is significantly suppressed.

[0101] The present invention also discloses an LSTM-based multi-radar system error registration device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned LSTM-based multi-radar system error registration method.

[0102] In another embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the various method embodiments described above.

[0103] Another embodiment of the present invention provides a computer program product that, when run on a data storage device, enables the data storage device to implement the steps in the various method embodiments described above.

[0104] If the integrated unit module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a storage device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0105] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0106] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0107] In the embodiments provided by this invention, it should be understood that the disclosed apparatus / device and method can be implemented in other ways. For example, the apparatus / device embodiments described above are merely illustrative. For instance, the division of modules or units is merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0108] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0109] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of this application.

Claims

1. A method for error registration in a multi-radar system based on LSTM, characterized in that, Includes the following steps: Acquire the target trajectory, which is located in the overlapping area of ​​multiple radar detection areas in a multi-radar system; Generate a feature sequence for each moment based on the target trajectory; Using the feature sequence at each time moment as input data, the location information of the target in the overlapping region at each time moment is determined by a probe region classification network; The system error of each radar in the multi-radar system is determined based on the location information. Generating a feature sequence for each time step based on the target trajectory includes: Generate the first target position coordinates at each moment in the multi-radar system coordinate system based on the target trajectory; Generate the second target position coordinates at each moment in the polar coordinate system based on the position coordinates of each radar in the multi-radar system and the target position coordinates; By combining the coordinates of the second target position corresponding to each radar at each time step, the feature sequence of each time step is obtained. The detection area classification network is an LSTM network, specifically including an LSTM layer, a fully connected layer, a softmax layer, and a classification layer connected in sequence. The input dimension of the LSTM layer is the same as the dimension of the feature sequence; Based on the location information, determine the system error of each radar in the multi-radar system: The sub-region number of the target in the overlapping area is determined based on the location information; Based on the sub-region number, query the system error table of each radar and determine the system error corresponding to the sub-region number in the system error table of each radar.

2. The LSTM-based multi-radar system error registration method as described in claim 1, characterized in that, The system error table is generated based on the radar's radial distance and azimuth angle.

3. The LSTM-based multi-radar system error registration method as described in claim 1 or 2, characterized in that, The training method for the LSTM network is as follows: Generate realistic target tracks; Different system errors are added to the real target track based on the system error table of each radar to obtain the detection track of each radar; Generate corresponding polar coordinate values ​​based on the detection tracks of each radar; By combining the polar coordinate values ​​corresponding to each radar at each time step, the training feature sequence at each time step is obtained; The training feature sequences at each time step are combined to generate the training sample set of the LSTM network.

4. The LSTM-based multi-radar system error registration method as described in claim 3, characterized in that, The loss function of the LSTM network is the cross-entropy classification loss function.

5. A multi-radar system error registration device based on LSTM, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements an LSTM-based multi-radar system error registration method as described in any one of claims 1 to 4.

Images