A radar track filtering method, device, electronic equipment, and storage medium

Abstract

This application belongs to the field of radar data processing technology, specifically disclosing a radar track filtering method, apparatus, electronic device, and storage medium. The method includes: acquiring current radar track observation data; inputting the radar track observation data into a radar track filtering model to obtain the current radar track filtering result output by the radar track filtering model; the radar track filtering model is trained based on radar track observation data samples and corresponding real track data labels, and is used to perform feature fusion based on sequence transformation features, time dependence features, and shape features obtained by feature extraction from the radar track observation data; and to determine the radar track filtering result based on the fused features. This application can accurately predict radar track data in complex scenarios, significantly improving the accuracy of radar track filtering in complex scenarios, and enhancing the precision and effectiveness of radar track filtering.

Description

Technical Field

[0001] This application belongs to the field of radar data processing technology, and more specifically, relates to a radar track filtering method, apparatus, electronic device and storage medium. Background Technology

[0002] Radar track filtering algorithms are a key technology in target tracking and navigation systems, aiming to eliminate measurement noise and improve the accuracy and continuity of target trajectories. The development of these filtering algorithms has roughly gone through three stages: early methods mainly relied on simple moving averages and polynomial fitting; subsequently, state-space model-based filtering methods became mainstream; and in recent years, with the improvement of computing power, machine learning methods have shown great potential in this field.

[0003] Existing deep learning-based track filtering methods typically use a single recurrent neural network (RNN), long short-term memory (LSTM), or gated recurrent unit (GRU) network. These methods extract a limited number of simple features and cannot accurately predict radar tracks in complex scenarios, resulting in low prediction accuracy and poor performance.

[0004] Therefore, how to better achieve radar track filtering has become a technical problem that the industry urgently needs to solve. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this application is to better realize radar track filtering, and to solve the problem of low prediction accuracy of existing track filtering methods in complex scenarios.

[0006] To achieve the above objectives, in a first aspect, this application provides a radar track filtering method, comprising: Acquire current radar track observation data; The radar track observation data is input into the radar track filtering model to obtain the current radar track filtering result output by the radar track filtering model. The radar track filtering model is trained based on radar track observation data samples and corresponding real track data labels. It is used to perform feature fusion based on the sequence transformation features, time dependence features and shape features obtained by feature extraction from the radar track observation data, and to determine the radar track filtering result based on the fused features.

[0007] Optionally, the radar track filtering model includes a data preprocessing module, a first feature extraction module, a second feature extraction module, a feature stitching module, and a track filter value calculation module; the output of the data preprocessing module is connected to the input of the first feature extraction module and the second feature extraction module, respectively; the outputs of the first feature extraction module and the second feature extraction module are connected to the input of the feature stitching module, and the output of the feature stitching module is connected to the input of the track filter value calculation module; The data preprocessing module is used to standardize the radar track observation data according to a preset standardization method to obtain standardized radar track observation data. The first feature extraction module is used to extract and fuse sequence transformation features from the standardized radar track observation data to obtain fused sequence transformation features; The second feature extraction module is used to extract time-dependent features and shape features from the standardized radar track observation data; The feature splicing module is used to splice the time-dependent features, the shape features, and the fused sequence transformation features to obtain the target fused features; The track filter value calculation module is used to determine the radar track filter result based on the target fusion features.

[0008] Optionally, the first feature extraction module includes a feature extraction sub-module and a multi-head attention mechanism module connected in sequence; The feature extraction submodule is built based on a TCN network and is used to extract the sequence transformation features of the standardized radar track observation data; The multi-head attention mechanism module is used to perform feature weight allocation and fusion processing on the sequence transformation features output by the feature extraction submodule through the multi-head attention mechanism to obtain the fused sequence transformation features.

[0009] Optionally, the track filter value calculation module includes a third feature extraction module, a fully connected calculation module, and an inverse normalization processing module connected in sequence; both the second feature extraction module and the third feature extraction module are constructed based on a ResGRU network. The third feature extraction module is used to extract time-dependent features and shape features from the target fusion features; The fully connected computation module is used to perform weighted computation on the output features of the third feature extraction module; The denormalization processing module is used to denormalize the calculation results output by the fully connected calculation module according to the preset normalization method to obtain the radar track filtering result.

[0010] Optionally, before inputting the radar track observation data into the radar track filtering model to obtain the current radar track filtering result output by the radar track filtering model, the method further includes: Radar track observation data samples and corresponding real track data labels are used as a set of training samples, and multiple sets of time-continuous training samples are obtained through a sliding window. The radar trajectory filtering model is trained using multiple sets of training samples.

[0011] Optionally, training the radar trajectory filtering model using the multiple sets of training samples includes: For any set of training samples, the training samples are input into the radar trajectory filtering model, and the predicted probability corresponding to the training samples is output. Using a preset loss function, the loss value is calculated based on the predicted probability corresponding to the training sample and the real track data label corresponding to the training sample; Based on the loss value, the model parameters of the radar trajectory filtering model are adjusted until the loss value is less than a preset threshold or the number of training iterations reaches a preset number. The model parameters obtained when the loss value is less than the preset threshold or when the number of training iterations reaches the preset number are used as the model parameters of the trained radar track filtering model, and the radar track filtering model training is completed.

[0012] Secondly, this application provides a radar track filtering device, comprising: The data acquisition module is used to acquire current radar track observation data; The track filtering module is used to input the radar track observation data into the radar track filtering model and obtain the current radar track filtering result output by the radar track filtering model. The radar track filtering model is trained based on radar track observation data samples and corresponding real track data labels. It is used to fuse sequence transformation features, time dependence features and shape features obtained by feature extraction from the radar track observation data, and to determine the radar track filtering result based on the fused features.

[0013] Thirdly, this application provides an electronic device, comprising: at least one memory for storing a program; and at least one processor for executing the program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to execute the method described in the first aspect or any possible implementation thereof.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0015] Fifthly, this application provides a computer program product that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0016] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0017] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: This application provides a radar track filtering method, apparatus, electronic device, and storage medium. By fully exploring the intrinsic correlation between high-dimensional time-series data features and real radar tracks in complex scenarios, the robustness and accuracy of the radar track filtering model are improved. Using the trained radar track filtering model, sequence transformation features, time dependence features, and shape features are extracted from radar track observation data with radar detection noise and fused together. Based on the fused features, real radar tracks are predicted. This method can accurately predict radar track data in complex scenarios, significantly improving the accuracy of radar track filtering in complex scenarios and enhancing the precision and effectiveness of radar track filtering. Attached Figure Description

[0018] Figure 1 This is one of the flowcharts illustrating the radar track filtering method provided in the embodiments of this application; Figure 2 This is a schematic diagram of a sliding window for capturing radar track observation data provided in an embodiment of this application; Figure 3 This is a schematic diagram of the radar track filtering model provided in the embodiments of this application; Figure 4 This is a second schematic flowchart of the radar track filtering method provided in the embodiments of this application; Figure 5 This is a schematic diagram of the TCN network structure provided in the embodiments of this application; Figure 6 This is a schematic diagram of the ResGRU network structure provided in the embodiments of this application; Figure 7 This is a schematic diagram of the radar trajectory filtering model training process provided in the embodiments of this application; Figure 8 This is a schematic diagram of the radar track filtering device provided in the embodiments of this application; Figure 9 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] The terms "first" and "second," etc., used in the specification and claims of this application are used to distinguish different objects, not to describe a specific order of objects. For example, "first feature extraction module" and "second feature extraction module," etc., are used to distinguish different feature extraction modules, not to describe a specific order of feature extraction modules.

[0021] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0022] In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple processing units means two or more processing units, multiple elements means two or more elements, etc.

[0023] The embodiments of this application are described below with reference to the accompanying drawings.

[0024] Figure 1 This is one of the flowcharts illustrating the radar track filtering method provided in the embodiments of this application, such as... Figure 1 As shown, it includes: Step S1: Obtain the current radar track observation data; Step S2: Input the radar track observation data into the radar track filtering model to obtain the current radar track filtering result output by the radar track filtering model; The radar track filtering model is trained based on radar track observation data samples and corresponding real track data labels. It is used to perform feature fusion based on the sequence transformation features, time dependence features and shape features obtained by feature extraction from radar track observation data, and to determine the radar track filtering result based on the fused features.

[0025] Specifically, the radar track observation data described in the embodiments of this application refers to the target track observation time series data generated by using observation radar to detect the target position. In practice, the observed track is not consistent with the actual track, and the observed track contains errors. Therefore, it is necessary to use the radar track observation data as input through track filtering to filter out its noise and obtain a filtered track that better fits the actual track.

[0026] It should be noted that regarding the format of radar track observation data, both the actual track and the observed track contain time and position information. However, because the radar processing system converts the radar's polar coordinate position to a Cartesian coordinate system, radar track observation data is represented as (time t, ...). x coordinate, y (Coordinates), whose zero point is the radar coordinates. Here, x coordinate, y Coordinates are used to correspond to the location information of the target being detected.

[0027] The radar track filtering model described in this application is obtained by training a preset deep neural network model based on radar track observation data samples and corresponding real track data labels. It is used to learn the intrinsic relationship between time series data features of different dimensions and real radar tracks. By extracting features from radar track observation data, sequence transformation features, time dependence features and shape features are obtained. Based on the features obtained by fusing these different dimensions of data features, radar track prediction is performed, and high-precision radar track filtering results are output.

[0028] Here, sequence transformation features refer to certain hidden features extracted from the original time series data through various linear or spatial transformations, which can more effectively characterize the inherent regularity and pattern of the sequence; time dependence features refer to features that are automatically learned and captured from the original time series data, which can reflect the dynamic patterns and dependencies of data changing over time; shape features are the most basic features in the time domain, and the numerical magnitude of the data sequence is precisely this feature.

[0029] It should be noted that, in the embodiments of this application, the radar track filtering model can be mainly constructed based on a deep neural network. Specifically, the deep neural network can be a neural network jointly constructed by a Temporal Convolutional Network (TCN) model and a Residual Gated Recurrent Unit (ResGRU) model, a neural network jointly constructed by a TCN network model and a Gated Recurrent Unit (GRU) model, or other neural networks that can be used to jointly extract sequence transformation features, time-dependent features, and shape features. No specific limitation is made in this application.

[0030] In this embodiment, the model training samples consist of multiple sets of radar track observation data samples carrying labels of real track data. It should be noted that the radar track observation data samples can be time-series data samples obtained by sequentially truncating the original radar track observation time-series data according to a sliding window of fixed length. It can be understood that the time-series length of each radar track observation data sample is equal to the sliding window length.

[0031] In embodiments of the present invention, the real track data label is predetermined based on radar track observation data samples and corresponds one-to-one with each radar track observation data sample. That is, each radar track observation data sample in the training samples is pre-set to carry a corresponding real track data label.

[0032] Furthermore, in the embodiments of this application, in step S1, the target's trajectory is detected by a preset radar detection device to obtain the current radar trajectory observation data. More specifically, in this step, the original dataset can be processed using a sliding window algorithm. The sliding window length is set to... The sliding step size is The starting point of the sliding window is Each three lines represent the target's trajectory at a single moment, with each line representing, in turn, the observed value. Observed values ,time Each column represents a time step. Time series data can be represented as , n This indicates the length of the data sequence. Therefore, by using a sliding window to extract data, the radar track observation data for the current window can be obtained. For example, as... Figure 2 As shown, the length of the sliding window is The sliding step size is The radar track observation data for the first window includes the first three columns of data, namely... .

[0033] Furthermore, in the embodiments of this application, in step S2, the current radar track observation data is input into a pre-trained radar track filtering model. The radar track filtering model can extract features from the radar track observation data to obtain corresponding sequence transformation features, time dependence features, and shape features. Then, these multi-dimensional data features are fused, and the current radar track filtering result is predicted based on the fused features.

[0034] The radar track filtering method in this application improves the robustness and accuracy of the radar track filtering model by fully exploring the intrinsic correlation between high-dimensional time-series data features and real radar tracks in complex scenarios. By using the trained radar track filtering model, sequence transformation features, time dependence features, and shape features are extracted from radar track observation data with radar detection noise and fused together. Based on the fused features, real radar tracks are predicted. This method can accurately predict radar track data in complex scenarios, significantly improving the accuracy of radar track filtering in complex scenarios and enhancing the precision and effectiveness of radar track filtering.

[0035] like Figure 3 and Figure 4 As shown, based on the above embodiments, as an optional embodiment, the radar track filtering model includes a data preprocessing module 1, a first feature extraction module 2, a second feature extraction module 3, a feature stitching module 4, and a track filter value calculation module 5; the output of the data preprocessing module 1 is connected to the input of the first feature extraction module 2 and the second feature extraction module 3 respectively; the outputs of the first feature extraction module 2 and the second feature extraction module 3 are connected to the input of the feature stitching module 4 respectively, and the output of the feature stitching module 4 is connected to the input of the track filter value calculation module 5; The data preprocessing module 1 is used to standardize the radar track observation data according to a preset standardization method to obtain standardized radar track observation data. The first feature extraction module 2 is used to extract and fuse sequence transformation features from standardized radar track observation data to obtain fused sequence transformation features. The second feature extraction module 3 is used to extract time-dependent features and shape features from standardized radar track observation data; Feature splicing module 4 is used to splice time-dependent features, shape features, and fused sequence transformation features to obtain target fused features; The track filter value calculation module 5 is used to determine the radar track filter result based on the target fusion features.

[0036] Specifically, the preset standardization method described in the embodiments of this application refers to a pre-set algorithm for data standardization processing, which may specifically employ the mean-variance standardization method. ).

[0037] In the embodiments of this application, a radar track filtering model is constructed by introducing a data preprocessing module, a first feature extraction module, a second feature extraction module, a feature stitching module, and a track filter value calculation module. The data preprocessing module is used to standardize the radar track observation data according to a preset standardization method to obtain standardized radar track observation data.

[0038] Optionally, in scenarios where outliers are not a concern, the pre-defined standardization method can employ max-min normalization to process the original dataset. Specifically, in scenarios sensitive to outliers, the method employs... The method employs data preprocessing to better mitigate the impact of outliers on radar track filtering accuracy. Specifically, this includes calculating the mean. The mean, representing the average of the entire dataset, describes the center of the dataset, and its calculation formula is: Then, calculate the standard deviation. Variance is the average of the squares of the differences between each data point and the mean, describing the dispersion of the dataset. The standard deviation is the arithmetic square root of the variance, and its formula is: Finally, calculate the standardized values. Standardize all numbers in the dataset to reduce the dataset to a smaller size. , In the distribution, The standardized formula is as follows: .

[0039] In the embodiments of this application, the first feature extraction module can be specifically constructed based on a deep neural network, used to extract and fuse sequence transformation features from the aforementioned standardized radar track observation data to obtain fused sequence transformation features.

[0040] like Figure 4 As shown, based on the above embodiments, as an optional embodiment, the first feature extraction module 2 includes a feature extraction sub-module 21 and a multi-head attention mechanism module 22 connected in sequence; The feature extraction submodule 21 is built based on the TCN network and is used to extract the sequence transformation features of standardized radar track observation data; The multi-head attention mechanism module 22 is used to perform feature weight allocation and fusion processing on the sequence transformation features output by the feature extraction submodule 21 through the multi-head attention mechanism to obtain the fused sequence transformation features.

[0041] Specifically, in the embodiments of this application, the first feature extraction module can be constructed using a feature extraction sub-module based on a TCN network and a multi-head attention (MHA) mechanism module. For example, Figure 4 The feature extraction submodule 21 in the network uses a TCN network structure. It's important to note that the difference between TCN and CNN lies in its convolution method: dilated causal convolution. Dilated causal convolution performs convolution operations in a skip-step manner, reducing redundant computations in ordinary convolution operations and exponentially expanding the receptive field (the receptive field refers to the range of features that can be extracted during each convolution operation). This allows TCN to achieve the same receptive field N as CNN with only LogN layers of dilated causal convolution.

[0042] Here, regarding the data processing method of the feature extraction submodule, such as... Figure 5 As shown, the convolution kernel of the TCN network is... Taking the matrix as an example, where The length of the convolution kernel, The width of the convolution kernel, The size is consistent with and remains unchanged in the feature dimension. The width of the convolution matrix is ​​such that the input for each convolution operation contains all features. It can be changed, and is generally an odd number greater than 3. The difference between causal convolution and ordinary convolution is that causal convolution pads the starting direction with zeros before the convolution operation, with a padding length of... This ensures that only past and current waypoints are used when calculating the convolution operation at each time step. The causal convolution formula is as follows: ; ; Perform dilated causal convolution: To determine the dilation factor, using the same size convolution kernel, the convolution operation on each column is increased during convolution calculation. The empty operation achieves an effect similar to expanding the convolution kernel, thereby increasing the receptive field. The formula for dilated causal convolution is as follows: ; ; Perform max pooling (Pool pooling) to pass features, and transfer the above pooled matrix. Divide into regions of equal size, that is And select the region. The largest value The eigenvalues ​​of this region are used to form a new feature matrix, which can be represented as: .

[0043] Activate the output. Use a ReLU (Renormalized Unified Network) layer to positively evaluate the output; the formula is as follows: ; Furthermore, a forgetting layer (Dropout) is used to forget parameters. During model training, a Dropout layer is periodically used to randomly disable some neurons to prevent the model from becoming overly reliant on certain neurons. The Dropout layer calculation steps are: randomly generate a mask, and the mask vector is... The length is the same as the number of parameters, and each element has a value of 0 or 1. During the forward propagation, the input... Use mask filtering .in, This indicates element-wise multiplication; when the mask corresponding to the parameter is 1, it means to retain; when the mask corresponding to the parameter is 0, it means to forget.

[0044] Furthermore, residual modules are used to connect the various networks. Let the input of the entire TCN network be... The output of the front-end network is The residual network output is , can be obtained Among them, the extracted sequence transformation features are .

[0045] like Figure 4 As shown, by The sequence transformation feature extraction process described above is executed repeatedly starting from =0. When, specify the length of a window. The data has been processed, and a window length has been obtained. The sequence transformation characteristics of radar track observation data at various times within the window can then be used to determine the window length. The observation data within will be used for the next stage of data processing.

[0046] It should be noted that this process will remove the original model from the learning process. Features transformed The characteristics of this feature not only help to better learn time-dependent features when using a ResGRU model, but also enable the construction of an identity network, transforming the output of a deep network being 0 into the weight parameters of the network being 0. Here, it is very difficult to get the output of a neural network to be 0, but it is easy to get the weight parameters of the network to be 0. This characteristic makes it possible to deepen the number of network layers.

[0047] Furthermore, in the embodiments of this application, the MHA mechanism in the multi-head attention mechanism module is used to re-divide the weights of the features using multiple sets of attention heads in parallel. Input matrix ( (For model dimensions). The calculation steps are as follows: Perform linear projection: per attention head (common Each (unit) learns an independent weight matrix, typically... , , , ,calculate , , The formula is as follows: , , ; Perform parallel computation: compute each head Calculate attention weights and output, using The function is normalized and a weight matrix is ​​generated. The calculation formula is as follows: ; Result fusion: The outputs of all attention heads are concatenated and subjected to a linear transformation. This process can be represented as: ; in, , This is the output weight matrix. Integrating all the formulas yields the fused sequence transformation features: .

[0048] The method in this embodiment constructs a first feature extraction module by combining a feature extraction sub-module based on a TCN network and an MHA mechanism module. The local features extracted by the TCN network provide a fine local context for MHA, while the global features extracted by the MHA mechanism supplement the TCN with long-range dependency information. The combination of the two can achieve seamless fusion of local and global features. At the same time, the model architecture is designed to use only past and current waypoints for calculation, preventing future information leakage and ensuring that calculations can be performed even when the input data is insufficient. This makes the model more adaptable and robust to different types of targets, maneuvering patterns, and data lengths.

[0049] Furthermore, in the embodiments of this application, the second feature extraction module can be specifically constructed based on a ResGRU network, used to extract time-dependent features and shape features from standardized radar track observation data.

[0050] It should be noted that analysis of the GRU network reveals limitations in its ability to extract time-dependent features. For excessively long time series, there is a risk of vanishing and exploding gradients. Although GRU has mitigated this risk to some extent, it still presents challenges with extremely long sequences. Therefore, introducing a residual model into GRU not only further addresses the risks of vanishing and exploding gradients but also transforms the features learned by the model from features at each time step to features derived directly from the differences between each time step. This makes it easier for the model to capture changes in the target data and strengthens ResGRU's ability to learn long-term dependent features.

[0051] Specifically, such as Figure 6 As shown, the ResGRU network introduces a residual module on top of the GRU network, transforming the network from learning features at each time step to learning features based on the differences between each time step. The total time step length is... The characteristics of the previous moment are The input features at the current time are The reset gate weight matrix is , bias is Update the gate weight matrix as follows , bias is The candidate state weight matrix is , bias is The calculation steps are as follows: First, calculate the reset gate, that is: ; The degree of impact of resetting gate control history information on the current candidate state, among which express The activation function is calculated using the following formula: ; Then, calculate the update gate, i.e.: ; Here, the update gate determines the fusion ratio of the old and new state information.

[0052] Then, the candidate hidden state is calculated, that is: ; The candidate hidden state is a temporary state that contains current information. Information after resetting the door ,in express product.

[0053] Finally, by incorporating residual modules during the computation of the final state through residual connections, the shape and time-dependent features of the time series can be extracted. This process can be represented as: ; .

[0054] Here, the shape feature of the time series is the specific numerical value of each element in the time-dependent feature matrix.

[0055] like Figure 4 As shown, by The above shape feature and time-dependent feature extraction process is executed repeatedly starting from =0. When, specify the length of a window. The data has been processed, and a window length has been obtained. The sequence transformation characteristics of radar track observation data at various times within the time frame.

[0056] Furthermore, in the embodiments of this application, the aforementioned time-dependent features, shape features, and fused sequence transformation features are spliced ​​together using a feature splicing module to obtain the target fused feature. Continuing to refer to... Figure 4 As shown, in addition to the TCN and MHA mechanisms on the left, the right side uses a ResGRU network, which, following the above processing method, yields shape features and time-dependent features. ,Right now Sequence variation features extracted from TCN network By splicing it together, the target fusion features can be achieved. ,Right now: ; Finally, by using the trajectory filter value calculation module in the radar trajectory filtering model, the real trajectory can be analyzed and predicted based on the target fusion features, and the final radar trajectory filtering result can be obtained.

[0057] The method in this application embodiment introduces a data preprocessing module, a first feature extraction module, a second feature extraction module, a feature splicing module, and a track filter value calculation module to jointly construct a radar track filtering model. It mines data features of radar track observation data from different dimensions and fuses them with the actual track prediction. This can effectively improve the robustness of the radar track filtering model in complex scenarios and improve the accuracy of radar track filtering.

[0058] Continue to refer to Figure 4 Based on the above embodiments, as an optional embodiment, the track filter value calculation module 5 includes a third feature extraction module 51, a fully connected calculation module 52, and an inverse normalization processing module 53 connected in sequence; both the second feature extraction module 3 and the third feature extraction module 51 are constructed based on a ResGRU network; The third feature extraction module 51 is used to extract time-dependent features and shape features from the target fusion features; The fully connected computation module 52 is used to perform weighted computation on the output features of the third feature extraction module; The denormalization processing module 53 is used to denormalize the calculation results output by the fully connected calculation module according to the preset normalization method to obtain the radar track filtering results.

[0059] Specifically, in the embodiments of this application, a third feature extraction module based on a ResGRU network, along with a fully connected computation module and an anti-normalization processing module, can be introduced to construct a track filter value calculation module. The third feature extraction module extracts time-dependent and shape features from the target fusion features.

[0060] Specifically, the target fusion feature values ​​obtained by the above splicing Further processing using a ResGRU network extracts time-dependent and shape features. =0 and the same shape feature and time-dependent feature extraction process described above will be executed repeatedly until... This indicates the length of a window. The data has been processed, allowing us to calculate the features corresponding to all time steps. The fully connected layer in the fully connected computation module is then used to activate the output, and the result is mapped to the real number domain. Here, it is assumed that a given input vector... Weight matrix bias The fully connected layer performs weighted calculations, and the result can be expressed as: ; Furthermore, in the embodiments of this application, the filtered values ​​are denormalized using a denormalization processing module. Since the input data was normalized during preprocessing, the output results need to be denormalized to obtain the true filtered values. As described above... The computational parameters of the method, let The denormalization processing module then performs denormalization calculations to obtain the radar track filtering results. It can be represented as: .

[0061] The method in this application constructs a trajectory filter value calculation module by introducing a third feature extraction module, a fully connected calculation module, and an anti-normalization processing module based on a ResGRU network. The TCN network structure, ResGRU network structure, and MHA mechanism used can all ensure that future information is not leaked, enabling the model to run even in extreme cases where the input data is insufficient. The dilated causal convolution used in the feature extraction process ensures that the model effectively extracts the relationship between features, thus ensuring the accuracy of the model in high-dimensional and complex scenarios.

[0062] Based on the above embodiments, as an optional embodiment, before step S2, in which radar track observation data is input into the radar track filtering model and the current radar track filtering result is obtained from the output of the radar track filtering model, the method further includes: Radar track observation data samples and corresponding real track data labels are used as a set of training samples, and multiple sets of time-continuous training samples are obtained through a sliding window. The radar trajectory filtering model is trained using multiple sets of training samples.

[0063] Specifically, in embodiments of the present invention, such as Figure 7As shown, before inputting radar track observation data into the radar track filtering model, the radar track filtering model needs to be trained to obtain a trained radar track filtering model.

[0064] It should be noted here that radar track observation data samples can be obtained by processing the original dataset using a sliding window algorithm. Similar to the parameter settings in the aforementioned application scenarios, the feature value length and label value length are set to [value missing]. (represents the time step used) The time step is obtained by filtering the trajectory of the step. (Filtered track), sliding window length is The sliding step size is The starting point of the sliding window is The original data length is Each set of 5 lines represents the target's trajectory at a single moment, with each line representing, in turn, the observation value. Observed values ,time The true value The true value Each column represents a time step. , No. i Data at a given time; time series data can be represented as Therefore, by using a sliding window to capture data, when... When, extract the starting column as Terminated 5 lines of Column data, window size is The calculation formula is as follows: ; when When the window has reached its boundary, the translation stops. The data captured by each sliding window is then concatenated to obtain multiple sets of time-continuous training sample datasets. This process can be represented as: .

[0065] Furthermore, all the sample data obtained above are divided into a training set and a test set. In this embodiment, the training set and the test set can be divided according to a ratio, for example, 7:3, that is, the radar track observation sample data of the first 70% of targets are extracted and stored in the training set, and the radar track observation sample data of the last 30% of targets are stored in the test set; then the radar track filtering model can be trained using the training set, and the performance of the trained radar track filtering model can be evaluated using the test set.

[0066] In an embodiment of the present invention, the radar track filtering model is trained using the aforementioned training set data. The specific training process is as follows: The radar track observation data sample and its corresponding real track data label are used as a set of training samples. That is, each radar track observation data sample with a real track data label is used as a set of training samples, and multiple sets of training samples are obtained in this way.

[0067] In embodiments of the present invention, radar track observation data samples correspond one-to-one with the actual track data tags they carry.

[0068] Then, after obtaining multiple sets of training samples, these sets are sequentially input into the radar trajectory filtering model to train the model. The radar track observation data samples and their accompanying real track data labels from each training sample group are simultaneously input into the radar track filtering model. For example... Figure 7 As shown, similar to the aforementioned data processing method, the radar track filtering model outputs the prediction results corresponding to each training iteration through the data preprocessing module, the first feature extraction module, the second feature extraction module, the feature stitching module, and the track filter value calculation module. It should be noted that the specific data processing flow for each module can be referred to the corresponding process in the aforementioned embodiments, and will not be elaborated here.

[0069] During the training process, each output result in the radar track filtering model is used to perform loss iterative calculation using a preset mean squared error (MSE) as the loss function. This iteratively adjusts the model parameters in the radar track filtering model until the preset training termination condition is met, thus completing the entire training process of the radar track filtering model and obtaining the trained radar track filtering model.

[0070] The method of this invention uses radar track observation data samples and their accompanying real track data labels as a set of training samples, and uses multiple sets of training samples to train the radar track filtering model, which helps to improve the model accuracy of the trained radar track filtering model.

[0071] Based on the above embodiments, as an optional embodiment, the radar trajectory filtering model is trained using multiple sets of training samples, including: For any set of training samples, the training samples are input into the radar track filtering model, and the predicted probability corresponding to the training samples is output. Using a preset loss function, the loss value is calculated based on the predicted probability corresponding to the training sample and the real track data label corresponding to the training sample; Based on the loss value, the model parameters of the radar trajectory filtering model are adjusted until the loss value is less than the preset threshold or the number of training iterations reaches the preset number. The model parameters obtained when the loss value is less than a preset threshold or the number of training iterations reaches a preset number are used as the model parameters of the trained radar track filtering model, and the radar track filtering model training is completed.

[0072] Specifically, the preset loss function described in the embodiments of the present invention refers to the loss function pre-set in the radar track filtering model for model evaluation; the preset threshold refers to the threshold pre-set in the model for obtaining the minimum loss value and completing model training; and the preset number of times refers to the maximum number of pre-set model iterations.

[0073] After obtaining multiple sets of training samples, for any set of training samples, the radar track observation data sample and its real track data label are simultaneously input into the radar track filtering model, and the predicted probability corresponding to the training sample is output.

[0074] Based on this, a preset loss function, such as MSE loss calculation, is used to calculate the loss value according to the predicted probability corresponding to the training sample and the real track data label corresponding to the training sample.

[0075] Furthermore, after calculating the loss value, the current training process ends. The model parameters of the entire radar track filtering model are then adjusted based on this loss value to update the weight parameters of each layer in the radar track filtering model. The next training iteration is then performed, and this process is repeated iteratively to train the model.

[0076] During training, if the training results for a certain set of training samples meet the preset training termination conditions, such as the calculated loss value being less than a preset threshold, or the model's loss value being controllable within the convergence range when the current iteration count reaches a preset number, then the model training ends. At this point, the obtained model parameters can be used as the model parameters of the trained radar track filtering model, thus completing the radar track filtering model training and obtaining the trained radar track filtering model.

[0077] The method of this invention iteratively trains the radar track filtering model by using multiple sets of time-continuous training samples captured by a sliding window, thereby controlling the loss value of the radar track filtering model within the convergence range, which helps to improve the accuracy of the radar track filtering results output by the model and improve the accuracy of radar track prediction.

[0078] The radar track filtering device provided in this application is described below. The radar track filtering device described below can be referred to in correspondence with the radar track filtering method described above.

[0079] Figure 8 This is a schematic diagram of the radar track filtering device provided in the embodiments of this application, as shown below. Figure 7 As shown, the device includes: Data acquisition module 10 is used to acquire current radar track observation data; The track filtering module 20 is used to input radar track observation data into the radar track filtering model and obtain the current radar track filtering result output by the radar track filtering model. The radar track filtering model is trained based on radar track observation data samples and corresponding real track data labels. It is used to fuse sequence transformation features, time dependence features and shape features obtained by feature extraction from radar track observation data, and to determine the radar track filtering result based on the fused features.

[0080] It is understood that the detailed functional implementation of each of the above units / modules can be found in the description in the aforementioned method embodiments, and will not be repeated here.

[0081] It should be understood that the above-described device is used to execute the methods in the above embodiments. The implementation principle and technical effect of the corresponding program modules in the device are similar to those described in the above methods. The working process of the device can be referred to the corresponding process in the above methods, and will not be repeated here.

[0082] The radar track filtering device of this application improves the robustness and accuracy of the radar track filtering model by fully exploring the inherent correlation between high-dimensional time-series data features and real radar tracks in complex scenarios. By using the trained radar track filtering model, sequence transformation features, time dependence features, and shape features are extracted from radar track observation data with radar detection noise and fused together. Based on the fused features, real radar tracks are predicted. This can accurately predict radar track data in complex scenarios, significantly improving the accuracy of radar track filtering in complex scenarios and enhancing the precision and effectiveness of radar track filtering.

[0083] Based on the methods in the above embodiments, this application provides an electronic device, such as... Figure 9 As shown, the electronic device may include a processor 910, a communications interface 920, a memory 930, and a communication bus 940, wherein the processor 910, the communications interface 920, and the memory 930 communicate with each other via the communication bus 940. The processor 910 can call logical instructions in the memory 930 to execute the methods in the above embodiments.

[0084] Furthermore, the logical instructions in the aforementioned memory 930 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0085] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0086] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0087] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0088] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0089] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0090] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0091] It should be understood that expressions such as “comprising” and “may include” used in this application indicate the existence of the disclosed functions, operations, or constituent elements, and do not limit one or more additional functions, operations, and constituent elements. In this application, terms such as “comprising” and / or “having” are to be interpreted as indicating a particular characteristic, number, operation, constituent element, component, or combination thereof, but not to exclude the existence or possibility of adding one or more other characteristics, numbers, operations, constituent elements, components, or combinations thereof.

[0092] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A radar track filtering method, characterized in that, include: Acquire current radar track observation data; The radar track observation data is input into the radar track filtering model to obtain the current radar track filtering result output by the radar track filtering model. The radar track filtering model is trained based on radar track observation data samples and corresponding real track data labels. It is used to perform feature fusion based on the sequence transformation features, time dependence features and shape features obtained by feature extraction from the radar track observation data, and to determine the radar track filtering result based on the fused features.

2. The radar track filtering method according to claim 1, characterized in that, The radar track filtering model includes a data preprocessing module, a first feature extraction module, a second feature extraction module, a feature stitching module, and a track filter value calculation module. The output of the data preprocessing module is connected to the input of the first feature extraction module and the second feature extraction module, respectively. The outputs of the first feature extraction module and the second feature extraction module are connected to the input of the feature stitching module, and the output of the feature stitching module is connected to the input of the track filter value calculation module. The data preprocessing module is used to standardize the radar track observation data according to a preset standardization method to obtain standardized radar track observation data. The first feature extraction module is used to extract and fuse sequence transformation features from the standardized radar track observation data to obtain fused sequence transformation features; The second feature extraction module is used to extract time-dependent features and shape features from the standardized radar track observation data; The feature splicing module is used to splice the time-dependent features, the shape features, and the fused sequence transformation features to obtain the target fused features; The track filter value calculation module is used to determine the radar track filter result based on the target fusion features.

3. The radar track filtering method according to claim 2, characterized in that, The first feature extraction module includes a feature extraction sub-module and a multi-head attention mechanism module connected in sequence; The feature extraction submodule is built based on a TCN network and is used to extract the sequence transformation features of the standardized radar track observation data; The multi-head attention mechanism module is used to perform feature weight allocation and fusion processing on the sequence transformation features output by the feature extraction submodule through the multi-head attention mechanism to obtain the fused sequence transformation features.

4. The radar track filtering method according to claim 2, characterized in that, The track filter value calculation module includes a third feature extraction module, a fully connected calculation module, and an inverse normalization processing module connected in sequence; both the second feature extraction module and the third feature extraction module are built based on a ResGRU network. The third feature extraction module is used to extract time-dependent features and shape features from the target fusion features; The fully connected computation module is used to perform weighted computation on the output features of the third feature extraction module; The denormalization processing module is used to denormalize the calculation results output by the fully connected calculation module according to the preset normalization method to obtain the radar track filtering result.

5. The radar track filtering method according to any one of claims 1-4, characterized in that, Before inputting the radar track observation data into the radar track filtering model to obtain the current radar track filtering result output by the radar track filtering model, the method further includes: Radar track observation data samples and corresponding real track data labels are used as a set of training samples, and multiple sets of time-continuous training samples are obtained through a sliding window. The radar trajectory filtering model is trained using multiple sets of training samples.

6. The radar track filtering method according to claim 5, characterized in that, The step of training the radar trajectory filtering model using the multiple sets of training samples includes: For any set of training samples, the training samples are input into the radar trajectory filtering model, and the predicted probability corresponding to the training samples is output. Using a preset loss function, the loss value is calculated based on the predicted probability corresponding to the training sample and the real track data label corresponding to the training sample; Based on the loss value, the model parameters of the radar trajectory filtering model are adjusted until the loss value is less than a preset threshold or the number of training iterations reaches a preset number. The model parameters obtained when the loss value is less than the preset threshold or when the number of training iterations reaches the preset number are used as the model parameters of the trained radar track filtering model, and the radar track filtering model training is completed.

7. A radar track filtering device, characterized in that, include: The data acquisition module is used to acquire current radar track observation data; The track filtering module is used to input the radar track observation data into the radar track filtering model and obtain the current radar track filtering result output by the radar track filtering model. The radar track filtering model is trained based on radar track observation data samples and corresponding real track data labels. It is used to fuse sequence transformation features, time dependence features and shape features obtained by feature extraction from the radar track observation data, and to determine the radar track filtering result based on the fused features.

8. An electronic device, characterized in that, Includes memory and one or more processors; The memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions; The one or more processors invoke the computer instructions to cause the electronic device to perform the method as described in any one of claims 1-6.

9. A computer-readable storage medium comprising instructions, characterized in that: When the instructions are executed on an electronic device, the electronic device causes the electronic device to perform the method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program or instructions, characterized in that: When the computer program or instructions are run on an electronic device, the electronic device causes the electronic device to perform the method as described in any one of claims 1-6.

Images