A gate adaptive radar data association method and device

Abstract

The present invention discloses a gate adaptive radar data association method and device, which obtains one-step predicted position information of multiple targets and echo information of multiple sensors; generates a first position Euclidean distance feature tensor between targets and a second position Euclidean distance feature tensor between targets and echoes based on the one-step predicted position information and the echo information; determines a gate threshold parameter of the target based on the first position Euclidean distance feature tensor and the position Mahalanobis distance feature tensor between the target and the echo; determines an association result of each target in each sensor based on the position Mahalanobis distance feature tensor between the target and the echo, the second position Euclidean distance feature tensor, and the gate threshold parameter; and by adaptively adjusting the gate threshold parameter, the present invention can better perceive the environment of the current tracking scene without relying on prior information, thereby reducing the computational burden for subsequent data association and lowering the computational difficulty.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target tracking, and in particular relates to a gate adaptive radar data association method and device. Background Art

[0002] With the continuous advancement of science and technology, people are no longer satisfied with target tracking using only a single sensor. Compared to single-sensor tracking systems, the core of multi-sensor tracking systems is to utilize measurement information from multiple sensors to improve target tracking stability and observation accuracy. To realize this advantage, it is first necessary to clarify the source of each measurement information, that is, to achieve data association.

[0003] However, in multi-target tracking, especially in dense clutter scenarios, the use of multi-sensor systems generates greater data volumes, posing even greater challenges to data association. On the one hand, this large volume of data challenges the algorithm's computational complexity and speed. On the other hand, achieving high-precision data association from multiple sensors challenges the algorithm's ability to process complex information and logic.

[0004] Traditional data association algorithms rely on prior information such as clutter density and distribution to achieve accurate association. When this prior information is inaccurate, algorithm performance is significantly impacted. In real-world applications, this information can be difficult or even unavailable. Furthermore, traditional algorithms are more suitable for scenarios with dispersed targets and sparse environmental clutter. When clutter is dense and targets are close together, the algorithm's association accuracy is low and can lead to combinatorial explosion and a surge in computational complexity. Summary of the Invention

[0005] The object of the present invention is to provide a gate adaptive radar data association method and device to reduce the amount of calculation when clutter is dense and the target is approaching.

[0006] The present invention adopts the following technical solution: a gate adaptive radar data association method, comprising the following steps:

[0007] Obtain one-step predicted position information of multiple targets and echo information of multiple sensors;

[0008] Generate a first position Euclidean distance feature tensor between targets, a second position Euclidean distance feature tensor between targets and echoes, and a position Mahalanobis distance feature tensor between targets and echoes according to the one-step predicted position information and the echo information;

[0009] Determine a gate threshold parameter of the target based on a first position Euclidean distance feature tensor and a position Mahalanobis distance feature tensor between the target and the echo;

[0010] The association result of each target in each sensor is determined according to the position Mahalanobis distance feature tensor between the target and the echo, the second position Euclidean distance feature tensor and the gate threshold parameter.

[0011] Furthermore, the gate threshold parameters are obtained through a CNN-based adaptive gate network module. The specific method includes:

[0012] Extract the feature information of the first position Euclidean distance feature tensor to obtain the inter-target feature information;

[0013] Extract the feature information of the position Mahalanobis distance feature tensor between the target and the echo to obtain the feature information between the target echoes;

[0014] splicing the inter-target feature information and the inter-target echo feature information to obtain spliced ​​feature information;

[0015] The splicing feature information is classified to obtain a category corresponding to the gate threshold parameter of each target; wherein the category is a size category of the gate threshold parameter.

[0016] Furthermore, the feature information of the position Mahalanobis distance feature tensor between the target and the echo is extracted, including:

[0017] The position Mahalanobis distance feature tensor between the target and the echo is mapped to a high-dimensional space through the first fully connected layer;

[0018] The corresponding target echo feature vectors are extracted through multiple one-dimensional convolutional layers with different kernel sizes;

[0019] Splicing the feature vectors between multiple target echoes;

[0020] The second fully connected layer is used to perform feature fusion on the spliced ​​target echo feature vectors to obtain the target echo feature information.

[0021] Furthermore, extracting feature information of the first position Euclidean distance feature tensor includes:

[0022] The first position Euclidean distance feature tensor is mapped to a high-dimensional space through a third fully connected layer, a batch normalization layer, and a fourth fully connected layer connected in series;

[0023] The corresponding inter-target feature vectors are extracted through multiple one-dimensional convolutional layers with different kernel sizes;

[0024] Concatenate feature vectors between multiple targets;

[0025] The fifth fully connected layer is used to perform feature fusion on the spliced ​​inter-target feature vectors to obtain inter-target feature information.

[0026] Furthermore, determining the association result of each target in each sensor includes:

[0027] The association probability between each target and each echo in each sensor is determined using a Bi-LSTM-based multi-sensor multi-target data association network module.

[0028] Furthermore, the association probability between each target and each echo in each sensor is determined using a Bi-LSTM-based multi-sensor multi-target data association network module, including:

[0029] Selecting candidate echo information for each target from the echo information of each sensor;

[0030] Reset all element values ​​in the second position Euclidean distance feature tensor that do not correspond to the candidate echo information;

[0031] The reset second position Euclidean distance feature tensor is input into the Bi-LSTM layer to obtain high-dimensional features;

[0032] Input the high-dimensional features into the sixth fully connected layer to obtain fused features;

[0033] Input the fused features into the multi-channel feature construction layer to obtain multi-channel features; the number of channels in the multi-channel features is the number of sensors;

[0034] The multi-channel features are input into the connected multi-channel one-dimensional convolutional layer and multiple seventh fully connected layers to obtain the association probability matrix of each target and each sensor.

[0035] Furthermore, the association probability matrix of each target and each sensor includes multiple one-dimensional vectors, each one-dimensional vector represents the association probability of a target with all echo information in a sensor; wherein the one-dimensional vector also includes the association probability of a target with missed detection.

[0036] Furthermore, determining the association probability between each target and each echo in each sensor using a Bi-LSTM-based multi-sensor multi-target data association network module also includes:

[0037] In each one-dimensional vector, the echo information or missed detection corresponding to the maximum association probability is selected as the association result of the corresponding target.

[0038] Furthermore, when a target is associated with multiple echoes, one of the echoes is selected for association with the target;

[0039] When an echo is associated with multiple targets, select one of the targets for association for the echo.

[0040] Another technical solution of the present invention is a gate adaptive radar data association device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor implements the above method when executing the computer program.

[0041] The beneficial effects of the present invention are as follows: the present invention can more accurately filter out clutter by adaptively adjusting the gate threshold parameters, ensuring that the actual measurement falls into the gate while the clutter falling into the gate is as small as possible, and better perceive the environment of the current tracking scene without relying on prior information, thereby reducing the computational burden for subsequent data association and reducing the computational difficulty. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flow chart of a gate adaptive radar data association method according to an embodiment of the present invention;

[0043] Figure 2 1 is a working process diagram of the CNN-based adaptive gate network module in an embodiment of the present invention;

[0044] Figure 3 This is a working process diagram of a multi-sensor multi-target data association network module based on Bi-LSTM in an embodiment of the present invention;

[0045] Figure 4 A schematic diagram of a simulation scenario in a simulation verification embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of the correlation accuracy of each target when tracking different methods in the simulation verification embodiment of the present invention;

[0047] Figure 6 This is a schematic diagram of the correlation accuracy of each target under different clutter densities and different methods in the simulation verification embodiment of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] Among traditional data association algorithms, the decision-based Near Neighbor Data Association (NNDA) is the simplest, directly associating a target with the measurement closest to its predicted position within its associated gate. However, it struggles with single-target tracking in dense clutter situations or tracking multiple targets with similar trajectories. Probabilistic Data Association (PDA) algorithms based on the Bayesian formula, and the Joint Probability Data Association (JPDA) algorithm, proposed based on the PDA algorithm, weight the measurements that fall within the associated gate to obtain measurements associated with the target. These algorithms improve data association accuracy but rely on prior information. PDA is suitable for single-target or sparsely populated multi-target tracking, while JPDA can adapt to more complex situations. However, when there are a large number of targets and candidate echoes, its computational complexity is high, making real-time operation difficult.

[0050] Traditional data association algorithms rely on prior information such as clutter density and distribution to achieve accurate association. When this prior information is inaccurate, algorithm performance is significantly impacted. In real-world applications, this information can be difficult or even unknowable. Furthermore, traditional algorithms are more suitable for scenarios with dispersed targets and sparse environmental clutter. When clutter is dense and targets are close together, the algorithm's association accuracy is low and can lead to combinatorial explosion and a surge in computational complexity.

[0051] Deep learning data association methods utilize recurrent neural networks as their primary structure. By inputting target attribute information and radar observations at a single or multiple moments, or by constructing feature information from both, the network generates a vector representing the probability of association between the target and all measurements. Association determination is then performed to complete the final association. These methods are still in the exploratory stage and lack maturity or systematicity. They fail to consider the more complex and challenging real-world environments faced by engineering applications, and their applicability to diverse scenarios needs improvement.

[0052] There are also some disadvantages in introducing data-driven methods: First, these methods are designed for single sensors and cannot efficiently handle data association problems in multi-sensor tracking systems. Second, since all measurements are directly used for association, some measurements that obviously do not belong to the target are also used as candidate measurements, which adds unnecessary difficulty to the task. At the same time, when the amount of clutter is large, it will also bring huge waste of resources. In addition, most of these methods directly use the target state value, the measurement state value, and the Euclidean distance between the two as network inputs, so that their generalization ability is still insufficient. When the tracking scene changes significantly (such as a change in the target state distribution), the performance will deteriorate.

[0053] To overcome the above shortcomings, the present invention designs an intelligent data association method with adaptive gates. This method uses a CNN-based adaptive gate module to replace the constant gate or the rule-based adaptive gate to mine implicit potential features from the target and measurement data, so that while ensuring that the actual measurement falls into the gate, the number of echoes falling into the gate is as small as possible. This gate that changes with scene factors helps to deal with scenarios where prior knowledge is scarce. At the same time, the Bi-LSTM-based association module is used to capture and learn the mutual influence that may occur when associating the data of each target, that is, the competition between different targets for the same measurement, as well as the potential information contained in the multi-sensor data, to achieve more accurate target-measurement association. This avoids the high computational burden and low timeliness introduced by classic algorithms such as JPDA when considering competition between targets.

[0054] In order to solve the problem of greatly increased data volume, association difficulty and complexity caused by multiple sensors, dense clutter and close targets, a data-driven method mainly composed of adaptive gate network and data association network is adopted to complete the data association task. The processing process is as follows: Figure 1 As shown, it is achieved through the following steps.

[0055] Specifically, the present invention discloses a gate adaptive radar data association method, comprising the following steps: obtaining one-step predicted position information of multiple targets and echo information of multiple sensors; generating a first position Euclidean distance feature tensor between targets, a second position Euclidean distance feature tensor between targets and echoes, and a position Mahalanobis distance feature tensor between targets and echoes based on the one-step predicted position information and the echo information; determining a gate threshold parameter of the target based on the first position Euclidean distance feature tensor and the position Mahalanobis distance feature tensor between targets and echoes; and determining an association result of each target in each sensor based on the position Mahalanobis distance feature tensor between targets and echoes, the second position Euclidean distance feature tensor, and the gate threshold parameter.

[0056] The present invention can more accurately filter out clutter by adaptively adjusting the gate threshold parameters, ensuring that the actual measurement falls into the gate while the clutter falling into the gate is as small as possible. Without relying on prior information, the environment of the current tracking scene can be better perceived, the computational burden for subsequent data association can be reduced, and the computational difficulty can be reduced.

[0057] That is, this method first constructs features of the echo information and target track information available in the current frame to obtain the input tensor of the network model. Specifically, it constructs the target-target position Euclidean distance feature tensor (i.e., the first position Euclidean distance feature tensor between targets), the target-echo position Euclidean distance feature tensor (i.e., the second position Euclidean distance feature tensor between target and echo), and the target-echo position Mahalanobis distance feature that takes into account the standard deviation of measurement noise (i.e., the position Mahalanobis distance feature tensor between target and echo).

[0058] Then, the first position Euclidean distance feature tensor and the position Mahalanobis distance feature tensor between the target and the echo are input into the CNN-based adaptive wave gate network module to learn the threshold size that adapts to the current scene. The working process of the CNN-based adaptive wave gate network module is as follows: Figure 2 shown.

[0059] First, the input feature data is normalized to its maximum and minimum values, and then feature extraction is performed on the two input feature vectors. That is, the feature information of the first position Euclidean distance feature tensor is extracted to obtain the inter-target feature information, and the feature information of the position Mahalanobis distance feature tensor between the target and the echo is extracted to obtain the inter-target echo feature information.

[0060] To obtain more effective feature mapping, different feature extraction networks are designed for the two input types. The positional Mahalanobis distance feature tensor between the target and the echo is mapped to a high-dimensional space via a first fully connected layer. Multiple one-dimensional convolutional layers with varying kernel sizes are then used to extract the corresponding inter-target echo feature vectors. These inter-target echo feature vectors are then concatenated, and finally, feature fusion is performed via a second fully connected layer to obtain inter-target echo feature information.

[0061] For the first-position Euclidean distance feature tensor, the feature extraction network is transformed based on the aforementioned network. Specifically, the part before entering the convolution is changed from a single-layer fully connected layer structure to a third fully connected layer, a batch normalization layer, and a fourth fully connected layer structure.

[0062] That is to say, the first-position Euclidean distance feature tensor is mapped to a high-dimensional space through the third fully connected layer, batch normalization layer, and fourth fully connected layer connected in series; the corresponding inter-target feature vectors are extracted through multiple one-dimensional convolutional layers with different kernel sizes; the multiple inter-target feature vectors are spliced; and the spliced ​​inter-target feature vectors are feature fused through the fifth fully connected layer to obtain inter-target feature information.

[0063] After obtaining target echo inter-feature information and target inter-echo feature information, the inter-target feature information and the target echo inter-feature information are spliced ​​to obtain spliced ​​feature information; the spliced ​​feature information is then classified to obtain a category corresponding to the gate threshold parameter of each target; wherein the category is a size category of the gate threshold parameter.

[0064] In other words, the two features extracted from the input are concatenated and fed into a classifier consisting of fully connected layers to obtain the adaptive threshold value for each target. Here, solving the adaptive threshold is modeled as a classification problem. The output adaptive thresholds are of different threshold sizes, such as small (e.g., 1-4), relatively small (e.g., 4-9), relatively large (e.g., 9-16), and large (e.g., 16-25). The number of classification categories and specific values ​​are determined according to the actual application.

[0065] After obtaining the threshold size level of the adaptive wave gate, the second position Euclidean distance feature tensor, the position Mahalanobis distance feature tensor and the wave gate threshold parameters are input into the Bi-LSTM-based multi-sensor multi-target data association network module to solve the association probability matrix of each target and each sensor, that is, to determine the association probability of each target with each echo in each sensor.

[0066] The working process of the multi-sensor multi-target data association network module based on Bi-LSTM is as follows: Figure 3 As shown in the figure. First, gate filtering is performed by comparing the position Mahalanobis distance feature tensor with the gate threshold parameter of each target. This determines the candidate echo corresponding to each target from each sensor. Specifically, candidate echo information for each target is selected from the echo information of each sensor. For example, if the position Mahalanobis distance feature tensor value between the target and the echo is 25 and the gate threshold parameter is 20, since 25>20, this echo is rejected.

[0067] Then, the second-position Euclidean distance feature tensor is normalized to its maximum and minimum values, and the normalized second-position Euclidean distance feature tensor is obtained based on the result of filtering out clutter with the wave gate. The specific operation is as follows: the element values ​​corresponding to the candidate echo in the normalized second-position Euclidean distance feature tensor remain unchanged, and the values ​​of the remaining elements are reset to 1, that is, the values ​​of the elements in the second-position Euclidean distance feature tensor that do not correspond to the candidate echo information are reset.

[0068] Next, the reset second-position Euclidean distance feature tensor is input into a deep learning network, which includes a Bi-LSTM layer, a fully connected layer, a multi-channel feature construction layer, a multi-channel one-dimensional convolutional layer, and multiple fully connected layers.

[0069] The Bi-LSTM layer considers the mutual influence of different target data when they are associated, taps into the potential information contained in multi-sensor data, and maps low-dimensional features into high-dimensional features. In other words, the reset second-position Euclidean distance feature tensor is input into the Bi-LSTM layer to generate high-dimensional features.

[0070] The fully connected layer fuses the different features extracted by the forward LSTM and backward LSTM in the Bi-LSTM layer, that is, the high-dimensional features are input into the sixth fully connected layer to obtain the fused features.

[0071] The multi-channel feature construction layer constructs the fused features into multi-channel features, where the number of channels is the number of sensors; that is, the fused features are input into the multi-channel feature construction layer to obtain multi-channel features; wherein the number of channels in the multi-channel features is the number of sensors.

[0072] Multi-channel one-dimensional convolutional layers extract and compress features. Finally, multiple fully connected layers map these features into a per-target, per-sensor association probability matrix. Each one-dimensional vector represents the probability distribution vector for the association between a target track and a sensor echo. In other words, multi-channel features are fed into a connected multi-channel one-dimensional convolutional layer and multiple seventh fully connected layers to produce a per-target, per-sensor association probability matrix.

[0073] Furthermore, the association probability matrix of each target and each sensor includes multiple one-dimensional vectors, each one-dimensional vector represents the association probability of a target with all echo information in a sensor; wherein the one-dimensional vector also includes the association probability of a target with missed detection.

[0074] Taking into account the possible missed detection, when calculating the probability distribution vector, in addition to calculating a corresponding association probability for each echo, a corresponding probability must also be calculated for the missed detection situation, that is, the probability that the target has no corresponding real measurement at the current moment and the current sensor.

[0075] Finally, the association matrix is ​​calculated through association judgment to determine the association results of each target for each sensor. Intuitively, the echo or missed detection with the highest probability of association with the target is selected as the target association result. That is, in each one-dimensional vector, the echo information or missed detection corresponding to the maximum association probability is selected as the corresponding target association result.

[0076] However, this cannot be done simply when a matching conflict occurs (for example, one echo is associated with multiple targets). This is because in the multi-target tracking process, each target track is associated with at most one echo from each sensor at each moment, and each echo can only be associated with at most one target.

[0077] Therefore, when solving the correlation results, we must first determine whether the above constraints are met by directly selecting the echo with the highest correlation probability as the correlation echo. If so, this is the final correlation result. Otherwise, if the constraints are not met, the echo correlation result is adjusted based on the principle of prioritizing the assignment of the echo with the highest correlation probability to the target with the smallest track number.

[0078] That is, when a target is associated with multiple echoes, one of the echoes is selected for association with the target; when an echo is associated with multiple targets, one of the targets is selected for association with the echo.

[0079] Simulation verification example:

[0080] Consider multi-source radar multi-target tracking in a two-dimensional rectangular coordinate system o-ξη. Here, the number of radars M = 3 and the sampling interval is 1s. After coordinate transformation, each radar (i.e., sensor) is located at the origin of the coordinate system, and the measurement matrix at time k in the corresponding measurement model is Where I represents the identity matrix, and the measurement noise covariance matrix R at time k is k =15 2 I2. At the same time, considering the dense clutter scene, the clutter is evenly distributed in the detection field of view, and the clutter density is λ = 10 -3 , and the number of clutter obeys the Poisson distribution with parameter λ.

[0081] The number of targets is T=5, and all targets move in a uniform linear motion in different directions within the observation range. In order to better display the target tracking results and compare the tracking performance of different algorithms, all measurements are uniformly displayed using the Cartesian coordinate system through coordinate transformation. Figure 4 As shown in Figure 1, targets 1, 2, and 3 can be considered as a formation, and targets 4 and 5 can be considered as a formation. The targets in the formation always have similar and close motion trajectories. The two formations approach each other between 10 and 20 seconds. The initial positions and motion states of the targets are shown in Table 1.

[0082] Table 1 Example of target initial position and motion state setting

[0083] Goal 1 Goal 2 Goal 3 Goal 4 Goal 5 X / m 700 700 700 75 100 VX / (m / s) -20 -20 -20 20 20 Y / m 1110 1050 1000 1150 1150 VY / (m / s) 0 0 0 -7 -7

[0084] The data association accuracy is designed as the evaluation indicator. The data association accuracy is defined as follows: For target i, its association accuracy is calculated as follows:

[0085]

[0086] Where k represents the sampling time, represents the probability of association between target i and the correct echo point of radar j in the tth test of the corresponding algorithm, Represents the total number of tests (i.e. the number of simulation scenarios).

[0087] It is worth noting that for the data association method of directly selecting the associated echo from the echo, such as the data association method and NNDA algorithm disclosed in the present invention, it is believed that if the selected measurement is the target true measurement, then is 1, otherwise it is 0. For data association methods that use the combined measurement obtained by weighted combination of candidate echoes as the associated echo, such as the PDA algorithm and the JPDA algorithm, it is considered that is the weighted probability value corresponding to the target true measurement in the combined measurement.

[0088] In order to study the performance of the method of the present invention, it was experimentally compared with the classical data association method and the method using the network. NNDA, PDA, and JPDA were selected as references for the classical data association method. Since there is no public neural network method for multi-sensors, a method based on the proposed multi-sensor multi-target data association network module but without considering gate filtering (i.e., DANet) and a method based on the proposed multi-sensor multi-target data association network module only considering fixed gate filtering (FGDANet) were selected as references. Here, Kalman filtering is used to achieve target track updates, and the above methods are respectively abbreviated as NNDA, PDA, JPDA, DANet, FGDANet and AGDAN (the method proposed in the present invention). At the same time, in order to clearly compare the tracking results of different algorithms, no track termination measures were taken.

[0089] Figure 5 The data association accuracy for each target is shown using different methods. Clearly, AGDAN achieves the highest association accuracy for each target (the horizontal and vertical axes in the figure represent k). For any target, the three classical algorithms struggle to achieve an association accuracy exceeding 60%, while the three intelligent methods all achieve an association accuracy exceeding 70%. This demonstrates that under the experimental conditions, the intelligent methods offer superior performance. This is because, compared to classical methods, they effectively integrate information from multiple sensors to aid in association.

[0090] Furthermore, it can be seen that as the two formations gradually approached and the distance between targets decreased, the association accuracy of all targets fluctuated between 10 and 20 seconds, with targets 2, 4, and 5 experiencing a significant decline. The ADGAN algorithm was minimally affected during this process, demonstrating the effectiveness of the adaptive gate network module. This module can capture changes in the relative positions of targets and, based on the current tracking status, adaptively adjust the size of the correlation gate to minimize clutter interference and narrow the range of candidate echoes during association, making data association accuracy less sensitive to environmental changes.

[0091] To further evaluate the algorithm performance, ADGAN is compared with other methods under different clutter density conditions. The average ADA (AADA) is selected as the evaluation metric. AADA represents the average value of the association accuracy of the five targets during the entire tracking process. Figure 6 As shown, the clutter density is λ=10 -5 and λ=5×10 -3 In almost all clutter density scenarios, the intelligent method significantly outperforms the classical method. In the intelligent method, except for λ = 5 × 10 -5 Except for this specific scenario, AGDAN achieves the highest accuracy; in this specific scenario, AGDAN's performance lags only slightly behind the state-of-the-art. Furthermore, AGDAN's AADA varies minimally with changes in scene clutter density, maintaining a consistently high level of performance. In summary, these results and analysis strongly demonstrate that the proposed method offers superior association accuracy and robustness.

[0092] In summary, the present invention designs a CNN-based adaptive gate network module that uses the target-target position Euclidean distance feature tensor and the target-echo position Euclidean distance feature tensor as input to solve the adaptive gate threshold for each target in the current frame. This gate threshold adaptively changes based on the relative spatial position relationship between targets and the spatial distribution of sensor echo positions. In this way, when filtering out clutter and determining candidate echoes through the gate, the amount of clutter that falls into the gate is minimized while ensuring that the actual measurement falls within the gate. Without relying on prior information, the environment of the current tracking scene can be better perceived, reducing the computational burden and difficulty for subsequent data association.

[0093] Furthermore, the present invention designs a multi-sensor, multi-target data association network module based on Bi-LSTM. This module uses the target-echo position Euclidean distance feature tensor, the target-echo position Mahalanobis distance feature tensor, and an adaptive gate threshold as input to calculate the association probability matrix between each target and each sensor echo. This module calculates association probabilities by exploiting the mutual influence of different target data associations and the potential information hidden in multi-sensor echo data, thereby improving the accuracy of the association results. Furthermore, when solving the association problem, it takes into account missed detections, thus improving the data association results.

[0094] Specifically, a CNN-based adaptive gate network module calculates the threshold for each target. This threshold changes as the target and echo information in the scene changes, minimizing the presence of clutter while still retaining true measurements. Echoes are filtered according to the gate threshold and fed into a Bi-LSTM-based data association network module, which leverages detection information from each sensor to calculate the association probability matrix.

[0095] As can be seen, the method presented here applies deep neural networks to the framework of classic data association methods. Compared to traditional methods, this reduces reliance on prior knowledge and avoids combinatorial explosion. Furthermore, compared to data-driven methods, it constrains the range of candidate echoes and improves association accuracy. This method achieves excellent tracking results while maintaining high association accuracy and reducing time consumption.

[0096] The present invention also discloses a gate adaptive radar data association device, which includes a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the above method is implemented.

[0097] The present invention further discloses an embodiment that provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned various method embodiments are implemented.

[0098] The present invention also provides a computer program product, which, when executed on a data storage device, enables the data storage device to implement the steps in the above-mentioned method embodiments.

[0099] If the integrated unit module is implemented in the form of 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, the present invention implements all or part of the process in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program, when executed by the processor, can implement the steps of the above-mentioned various method embodiments. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include at least: any entity or device capable of carrying the computer program code to a storage device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electric carrier signal, a telecommunication signal and a software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk.

[0100] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0101] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the present invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0102] In the embodiments provided herein, it should be understood that the disclosed devices and methods may be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of modules or units is merely a logical functional division. In actual implementation, other division methods may be used, such as combining or integrating multiple modules or components into another system, or omitting or not implementing certain features.

[0103] The modules described as separate components may or may not be physically separate, and 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 elements. Some or all of the modules may be selected to achieve the purpose of the solution of this embodiment according to actual needs.

Claims

1. A gate adaptive radar data association method, characterized in that: The following steps are involved: Obtain one-step predicted position information of multiple targets and echo information of multiple sensors; generating a first position Euclidean distance feature tensor between targets, a second position Euclidean distance feature tensor between targets and echoes, and a position Mahalanobis distance feature tensor between targets and echoes according to the one-step predicted position information and the echo information; Obtaining a gate threshold parameter based on the first position Euclidean distance feature tensor and the position Mahalanobis distance feature tensor between the target and the echo through a CNN adaptive gate network module; The association result of each target in each sensor is determined by a Bi-LSTM multi-sensor multi-target data association network module according to the position Mahalanobis distance feature tensor between the target and the echo, the second position Euclidean distance feature tensor and the wave gate threshold parameter.

2. The gate adaptive radar data association method according to claim 1, wherein: The gate threshold parameters are obtained by a CNN-based adaptive gate network module, specifically by: Extracting feature information of the Euclidean distance feature tensor at the first position to obtain inter-target feature information; Extract the feature information of the position Mahalanobis distance feature tensor between the target and the echo to obtain the feature information between the target echoes; splicing the inter-target feature information and the inter-target echo feature information to obtain spliced ​​feature information; The splicing feature information is classified to obtain a category corresponding to the gate threshold parameter of each target; wherein the category is a size category of the gate threshold parameter.

3. The gate adaptive radar data association method according to claim 2, wherein: Extracting feature information of the position Mahalanobis distance feature tensor between the target and the echo includes: The position Mahalanobis distance feature tensor between the target and the echo is mapped to a high-dimensional space through the first fully connected layer; The corresponding target echo feature vectors are extracted through multiple one-dimensional convolutional layers with different kernel sizes; splicing the characteristic vectors between a plurality of target echoes; The second fully connected layer performs feature fusion on the spliced ​​target echo feature vectors to obtain the target echo feature information.

4. The gate adaptive radar data association method according to claim 3, wherein: Extracting feature information of the Euclidean distance feature tensor of the first position includes: Mapping the first position Euclidean distance feature tensor to a high-dimensional space through a third fully connected layer, a batch normalization layer, and a fourth fully connected layer connected in series; The corresponding inter-target feature vectors are extracted through multiple one-dimensional convolutional layers with different kernel sizes; splicing a plurality of said inter-target feature vectors; The fifth fully connected layer performs feature fusion on the spliced ​​inter-object feature vectors to obtain the inter-object feature information.

5. A gate adaptive radar data association method according to any one of claims 2 to 4, characterized in that: Determining the association results for each target in each sensor includes: The association probability between each target and each echo in each sensor is determined using a Bi-LSTM-based multi-sensor multi-target data association network module.

6. The gate adaptive radar data association method according to claim 5, characterized in that: The Bi-LSTM-based multi-sensor multi-target data association network module is used to determine the association probability between each target and each echo in each sensor, including: Selecting candidate echo information for each target from the echo information of each sensor; Resetting all element values ​​in the second-position Euclidean distance feature tensor that do not correspond to the candidate echo information; The reset second position Euclidean distance feature tensor is input into the Bi-LSTM layer to obtain high-dimensional features; Inputting the high-dimensional features into the sixth fully connected layer to obtain fused features; Inputting the fused features into a multi-channel feature construction layer to obtain a multi-channel feature; wherein the number of channels in the multi-channel feature is the number of sensors; The multi-channel features are input into the connected multi-channel one-dimensional convolutional layers and multiple seventh fully connected layers to obtain the association probability matrix of each target and each sensor.

7. The gate adaptive radar data association method according to claim 6, characterized in that: The target-sensor association probability matrix includes multiple one-dimensional vectors, each of which represents the association probability of a target and all echo information of a sensor; wherein the one-dimensional vector also includes the association probability of a target and a missed detection situation.

8. The gate adaptive radar data association method according to claim 7, wherein: The Bi-LSTM-based multi-sensor multi-target data association network module is used to determine the association probability between each target and each echo in each sensor, and further includes: In each one-dimensional vector, the echo information or missed detection corresponding to the maximum association probability is selected as the association result of the corresponding target.

9. The gate adaptive radar data association method according to claim 8, characterized in that: When a target is associated with multiple echoes, one of the echoes is selected for association for the target; When an echo is associated with multiple targets, select one of the targets for association for the echo.

10. A gate adaptive radar data association device, 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, the method according to any one of claims 1 to 9 is implemented.

Images