A high-maneuverability target tracking method based on frequency domain modeling and denoising network

Abstract

The present application provides a high-maneuverability target tracking method based on a frequency domain modeling and denoising network, which relates to the field of digital signal processing, including: obtaining an observation trajectory sequence of a high-maneuverability target; preprocessing the observation trajectory sequence to obtain a preprocessing sequence; inputting the preprocessing sequence into an adaptive frequency domain denoising tracking network for tracking filtering processing to obtain a frequency domain filtering result; wherein the adaptive frequency domain denoising tracking network is used to filter out noise, and extract the global features of the preprocessing sequence in the frequency domain, and extract the local features of the preprocessing sequence in the time domain; and tracking the high-maneuverability target according to the frequency domain filtering result. The present application can achieve accurate modeling of the target maneuvering time domain, thereby improving tracking accuracy, avoiding the large amount of calculations brought by the attention mechanism, and improving computational efficiency.

Description

Technical Field

[0001] The present application relates to the field of digital signal processing, and in particular to a high-maneuverability target tracking method based on a frequency domain modeling and noise reduction network. Background Art

[0002] With the rapid development of information technology, target tracking technology has been widely used in many fields such as drone monitoring, surveillance security, unmanned driving, human-computer interaction, etc. Target tracking estimates the motion state of the target through sensor measurement data (such as radar, infrared or sonar). Target tracking can usually be divided into maneuvering target tracking and non-maneuvering target tracking, where "maneuvering" means that the target suddenly changes its motion state when it is unpredictable. Maneuvering target tracking is to estimate the high-dimensional sequence composed of the actual position, velocity and acceleration of the target based on the received target observation sequence. This problem has always been very challenging. The challenge comes from two aspects. First, due to limited observation conditions, the target position deviates from the true position of the target due to interference such as line of sight accuracy and background clutter; second, the target itself moves maneuverably, making it difficult to infer the high-dimensional state parameters such as the target's velocity and acceleration through linear methods. At present, the methods to solve this problem are mainly divided into two categories, one is the traditional filtering algorithm based on the model, and the other is the intelligent algorithm based on deep learning.

[0003] For high-maneuver target tracking based on frequency domain modeling and denoising networks, traditional methods use multiple models to fit complex motion states, i.e., multi-model algorithms. In order to achieve better synergy between multiple models, interactive multiple model algorithms are widely used, which greatly improves the performance of high-maneuver target tracking based on frequency domain modeling and denoising networks. Therefore, IMM (Interacting Multiple Model) has gradually become a widely used tracking framework. However, model-based methods need to accumulate some observations to determine some parameters of the model, which leads to delayed estimation problems. Summary of the invention

[0004] The purpose of this application is to provide a high-maneuverability target tracking method based on a frequency domain modeling and denoising network, which captures the global information of the target trajectory in the frequency domain and adaptively filters out noise to solve the problem of low efficiency and poor accuracy of existing target tracking.

[0005] In order to solve the above technical problems, the present application provides a high-maneuverability target tracking method based on frequency domain modeling and denoising network. The specific technical solution is as follows:

[0006] Obtaining the observed trajectory sequence of highly maneuverable targets;

[0007] Preprocessing the observation trajectory sequence to obtain a preprocessed sequence;

[0008] The preprocessing sequence is input into an adaptive frequency domain noise reduction tracking network for tracking and filtering processing to obtain a frequency domain filtering result; wherein the adaptive frequency domain noise reduction tracking network is used to filter out noise, and extract the global features of the preprocessing sequence in the frequency domain, and extract the local features of the preprocessing sequence in the time domain;

[0009] The high-maneuverability target is tracked according to the frequency domain filtering result.

[0010] Optionally, the preprocessing of the observation trajectory sequence includes:

[0011] Performing coordinate system transformation on the observation trajectory sequence;

[0012] Determining the velocity components of the high maneuverability target so that the input sequence feature dimension and the output dimension are consistent; and determining that the components of each dimension conform to a normal distribution;

[0013] Normalization processing is performed on different physical quantities of the observation trajectory sequence.

[0014] Optionally, the step of inputting the preprocessed sequence into an adaptive frequency domain noise reduction tracking network for tracking filtering includes:

[0015] In the adaptive frequency domain denoising tracking network, a timestamp component is added to each of the preprocessed sequences using an input layer, and the preprocessed sequences with the timestamp component added are mapped to a feature space through a linear layer;

[0016] After the preprocessing sequence is mapped, it is input into the cascaded filter layer for processing to obtain a frequency domain filtering result; wherein the filter layer includes a frequency domain processing layer and a time domain processing layer.

[0017] Optionally, the input cascade filter layer is processed including:

[0018] Transforming the preprocessed sequence into the frequency domain through a one-dimensional discrete Fourier transform;

[0019] In the frequency domain, one of the variables is filtered by a global filter to obtain a first filtering result; the other variable is first filtered in the frequency domain by an adaptive mask containing a power spectrum map and then input into the global filter to obtain a second filtering result;

[0020] The first filtering result and the second filtering result are superimposed to obtain the frequency domain filtering result.

[0021] Optionally, when the input is processed in the cascaded filter layer and the frequency domain filtering result is obtained, it also includes:

[0022] Before being input into the time domain processing layer, the frequency domain filtering result is converted into the time domain through a one-dimensional inverse discrete Fourier transform.

[0023] Optionally, the input is processed in the cascaded filter layer, and the frequency domain filtering results obtained include:

[0024] In the time domain processing layer, at least two convolutional layers with kernels of different sizes are used to extract time domain local features respectively;

[0025] The two-way time domain local features are fused as the supervisory signals of the other way respectively, and the final features are obtained after fusion;

[0026] The final feature is mapped back to the target real physical space to obtain a frequency domain filtering result.

[0027] Optionally, tracking the high-maneuverability target according to the frequency domain filtering result includes:

[0028] The frequency domain filtering result is generated to generate a filtering effect graph, and the high-maneuverability target is tracked according to the filtering effect graph.

[0029] Optionally, performing coordinate system conversion on the observation trajectory sequence includes:

[0030] The observation trajectory sequence is converted from a polar coordinate system to a plane coordinate system according to the observation trajectory sequence, the converted time, the distance of the target from the sensor, and the observation angle.

[0031] Optionally, determining the velocity component of the high maneuverability target includes:

[0032] The velocity component of the high maneuverable target is estimated using a time domain difference method; the time domain difference method is used to discretize the observed trajectory sequence in the time domain, use finite difference approximation to calculate the rate of change, and obtain an estimated value of the velocity component based on the rate of change.

[0033] Optionally, performing normalization processing on different physical quantities of the observation trajectory sequence includes:

[0034] A normalization formula is used to perform normalization processing on different physical quantities of the observation trajectory sequence.

[0035] The present application provides a high-maneuverability target tracking method based on a frequency domain modeling denoising network, comprising: obtaining an observation trajectory sequence of a high-maneuverability target; preprocessing the observation trajectory sequence to obtain a preprocessing sequence; inputting the preprocessing sequence into an adaptive frequency domain denoising tracking network for tracking filtering processing to obtain a frequency domain filtering result; wherein the adaptive frequency domain denoising tracking network is used to filter out noise, and extract global features of the preprocessing sequence in the frequency domain, and extract local features of the preprocessing sequence in the time domain; and tracking the high-maneuverability target according to the frequency domain filtering result.

[0036] The present application can capture the global information of the target maneuver in the frequency domain, thereby adaptively filtering out noise, and can also capture local features through local time domain pattern supervision, thereby achieving accurate modeling of the target maneuver in the time domain and improving tracking accuracy; and all operations utilize convolutional neural networks, avoiding the large amount of calculation brought by the attention mechanism, thereby improving computing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0038] Figure 1 A flowchart of a high-maneuverability target tracking method based on a frequency domain modeling denoising network provided in an embodiment of the present application;

[0039] Figure 2 A schematic diagram of the observation trajectory provided in the embodiment of the present application;

[0040] Figure 3 A schematic diagram of the frequency domain filter structure provided in an embodiment of the present application;

[0041] Figure 4 A schematic diagram of the processing process of the time domain processing module provided in an embodiment of the present application;

[0042] Figure 5 A schematic diagram of 100 high-maneuverability target trajectories in the same scenario provided by an embodiment of the present application;

[0043] Figure 6 A diagram showing the tracking trajectories of various methods in the first scenario provided in the embodiment of the present application;

[0044] Figure 7 A comparison chart of tracking performance of various methods in the first scenario provided in the embodiment of the present application;

[0045] Figure 8 A diagram showing the tracking trajectories of various methods in the second scenario provided in the embodiment of the present application;

[0046] Fig. 9 This is a comparison chart of the tracking performance of various methods in the second scenario provided in the embodiment of the present application. DETAILED DESCRIPTION

[0047] The purpose of the present invention is to solve the problem of low efficiency and poor precision of existing target tracking. Focusing on the application scenario of ranging and angle measurement system, an adaptive target tracking method based on frequency domain modeling and noise reduction is provided.

[0048] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0049] Please refer to Figure 1 , Figure 1 A flowchart of a high-maneuverability target tracking method based on a frequency domain modeling denoising network provided in an embodiment of the present application, the method comprising:

[0050] S101: Acquire the observation trajectory sequence of the high maneuverability target;

[0051] S102: preprocessing the observation trajectory sequence to obtain a preprocessed sequence;

[0052] S103: Input the preprocessed sequence into an adaptive frequency domain noise reduction tracking network for tracking and filtering processing to obtain a frequency domain filtering result; wherein the adaptive frequency domain noise reduction tracking network is used to filter out noise, and extract the global features of the preprocessed sequence in the frequency domain, and extract the local features of the preprocessed sequence in the time domain;

[0053] S104: Tracking the high-maneuverability target according to the frequency domain filtering result.

[0054] The observation model under the ranging and angle measurement system is shown as follows:

[0055] ;

[0056] ;

[0057] in, For the goal The horizontal and vertical coordinates of the rectangular coordinate system are always observed, and the noise will be affected during the observation process. (representing angle-related noise and distance-related noise, respectively) interfere, causing both the distance measurement and angle measurement information to deviate from the target's true position, and ultimately obtaining the target's angle measurement and distance measurement information The schematic diagram of the observation trajectory is as follows Figure 2 As shown, Figure 2 This is a schematic diagram of the observation trajectory provided in the embodiment of the present application. Figure 2 The left figure in the middle shows the angle measurement information changing with time, and the right figure shows the distance measurement information changing with time.

[0058] This application separately learns the frequency domain features and time domain features of the target observation sequence to enhance the neural network's ability to perceive the global and local features of the target trajectory sequence, so that the method can adaptively give high-precision filtering results when the target is maneuvering. The specific process is as follows:

[0059] The adaptive spectral denoising tracking network (ASDN) proposed by the present invention can be composed of an adaptive spectral denoising filter (ASDF) module and a local-pattern guided time domain module (LGTB). The adaptive spectral denoising tracking network can be composed of these two modules in cascade.

[0060] When the platform load remains unchanged, the noise at different times is unrelated white noise, while the target state at different times is related. The ASDF module learns the frequency domain feature distribution by frequency domain feature modeling to achieve the purpose of noise reduction. The LGTB module considers the variability of the target motion model and needs to enhance the network's perception of local features. Therefore, the design of dual-branch local feature mutual supervision is used to strengthen the guidance of local features.

[0061] First, the observation trajectory sequence of the high-maneuvering target is obtained. There is no limitation on how to obtain the observation trajectory sequence. It can be achieved by modeling the high-maneuvering target, sensor measurement and filter design. Among them, the modeling of the high-maneuvering target describes the change process of the target state over time, while the sensor measurement provides the necessary measurement data. Finally, the target state is estimated and predicted by filter design. When modeling the high-maneuvering target, the uniform motion CV model can be used: when the target performs uniform linear motion, the CV model is the most commonly used description method. This model assumes that the speed of the target remains constant in all directions, so it can be used to describe the motion of targets whose speed does not change much. Of course, other models can also be used: in addition to the CV model, there are also many models such as the uniform acceleration CA model, the uniform turning CT model, the Singer model, etc. The measurement data usually comes from detectors such as radar, infrared, sonar or electromagnetic sensors. These data include kinematic parameters such as the position and speed of the target, and may contain noise and clutter. In order to improve the signal-to-noise ratio, the measurement data can be preprocessed, which includes data compression, equal weight and variable weight preprocessing, and outlier removal techniques, aiming to optimize the data quality to improve the accuracy of subsequent steps. When performing filtering, a Kalman filter or a nonlinear filtering method may be used.

[0062] In addition, the observation trajectory sequence can also be derived from historical data.

[0063] After obtaining the observation trajectory sequence, the observation trajectory sequence of the high-maneuverability target can be preprocessed to improve the network learning and reasoning efficiency. Specifically, it can be divided into three steps:

[0064] Next, transform the observations from polar coordinates to planar coordinates:

[0065] Original observation sequence , at the time , the target distance to the sensor is , the observation angle is . Convert to plane coordinate system:

[0066] ;

[0067] ;

[0068] in, The time after conversion The image plane coordinates of .

[0069] The second step is to determine the velocity components of the high maneuverability target so that the input sequence feature dimension and the output dimension are consistent; and to determine that the components of each dimension conform to the normal distribution:

[0070] ;

[0071] ;

[0072] ;

[0073] in, for Estimation of the target image plane position and velocity at the moment, Respectively represent Always on target Axis and The estimation of the axis velocity, corresponding to It represents the estimated speed of the target at time 1.

[0074] As shown in the above formula, the velocity components of the target are roughly obtained based on the observation through time domain difference, so that the components of each dimension can conform to the normal distribution. .

[0075] The third step is to normalize the different physical dimensions of the observation sequence to a unified scale:

[0076] ;

[0077] ;

[0078] in, is the time mean of the observation sequence, is the normalized observed variable.

[0079] After S102 is completed, step S103 is executed, that is, the preprocessed sequence is sent to the adaptive frequency domain noise reduction tracking network for tracking filtering. Specifically, in the adaptive frequency domain noise reduction tracking network, the input layer is used to add a timestamp component to each of the preprocessed sequences, and the preprocessed sequence with the timestamp component added is mapped to the feature space through a linear layer. After the preprocessed sequence is mapped, it is input into the cascaded filter layer for processing to obtain a frequency domain filtering result; wherein, the filter layer includes a frequency domain processing layer and a time domain processing layer.

[0080] The input layer adds a timestamp component to each sequence, which is then mapped to the feature space through a linear layer.

[0081] After the mapping process of the input layer, it is sent to the cascade filter layer for processing. The filter layer is divided into frequency domain processing and time domain processing. Figure 3 , Figure 3 This is a schematic diagram of the frequency domain filter structure provided in the embodiment of the present application. Figure 3 It is the adaptive frequency domain filtering module ASDF. Figure 3 middle," " indicates pixel-level multiplication, " " means pixel-level addition. DFT stands for Discrete Fourier Transform. IDFT stands for Inverse Discrete Fourier Transform.

[0082] For frequency domain filters, the input is first transformed into the frequency domain via a one-dimensional discrete Fourier transform:

[0083] ;

[0084] The filter is then divided into two branches, one is directly filtered by the global filter, and the other is sent to the global filter before passing through a layer of adaptive mask obtained by power spectrum mapping for frequency domain filtering. The two variables are then added together to obtain the final frequency domain filtering result. It is worth noting that the global filter and adaptive mask mentioned here are parameters obtained by the network based on a large amount of training data, without the need for artificial prior settings. Before being input into the time domain processing layer, the frequency domain filtering result is converted to the time domain through a one-dimensional discrete Fourier inverse transform.

[0085] See also Figure 4 , Figure 4A schematic diagram of the processing process of the time domain processing module provided in an embodiment of the present application, Figure 4 In the above, LGTB represents the time domain processing module, and GELU is the activation function (Gaussian Error Linear Unit). In the time domain processing layer, at least two convolutional layers with kernels of different sizes are used to extract local time domain features. The two local time domain features are fused as the supervisory signals of the other channel, and the final features are obtained after fusion. The final features are mapped back to the target real physical space to obtain the frequency domain filtering result. It can be seen that the time domain processing module can be divided into two branches, containing two convolutional layers with kernels of different sizes, namely Figure 4 As shown in the figure, C1K3 and C1K1, C1K1 indicates that a convolution layer has 1 convolution kernel of size 1×1. C1K3 indicates that a convolution layer has 1 convolution kernel of size 3×3. That is, the convolution kernel size of a convolution layer is 1, and the size of a convolution kernel is 3. Through this design, the temporal local features of different receptive fields of the two branches are captured. Then the outputs of the two branches are fused as the supervision signals of the other branch respectively. At the end of the module, the features of the two branches are directly added. The final features are obtained.

[0086] After the cascaded filtering layer is completed, this application will map the filtered features back to the target real physical space to obtain the tracking result.

[0087] ;

[0088] In the above formula, is the estimated target state sequence component.

[0089] The embodiments of the present application can capture the global information of the target maneuver in the frequency domain, thereby adaptively filtering out noise, and can also capture local features through local time domain pattern supervision, thereby achieving accurate modeling of the target maneuver in the time domain and improving tracking accuracy; and all operations utilize convolutional neural networks, which avoids the large amount of calculation brought by the attention mechanism and improves computing efficiency.

[0090] The following is a description of a specific application scenario of this application:

[0091] This application sets the network specific feature dimension to 64. The discrete Fourier transform DFT and its inverse transform Inverse DFT (IDFT) in the specific implementation of the network are executed by the interface provided by the Pytorch library. This application sets the cascade filter layer to three layers, and the network Adam optimizer uses a linear decay learning rate of 1e-4 to schedule iterations for 100 periods, with a decay rate of 1e-6. The training data set is a high-mobility target trajectory data set publicly available on the network.

[0092] A visual verification is performed on two extremely high maneuverability target cases, and a statistical verification is performed on the trajectories of 3,000 different highly maneuverable targets from the same perspective.

[0093] 100 simulated high-maneuverability target trajectories are selected as an example. Figure 5 As shown, Figure 5 This is a schematic diagram of 100 high-maneuverability target trajectories in the same scenario provided in an embodiment of the present application.

[0094] The tracking results of this application and other leading methods in the current field for extremely high maneuvering targets are as follows: Figures 6 to 9 As shown. Among them, Figure 6 This is a diagram showing the tracking trajectories of various methods in the first scenario provided in the embodiment of the present application. Figure 7 This is a comparison chart of the tracking performance of various methods in the first scenario provided in the embodiment of the present application. Figure 8 This is a diagram showing the tracking trajectories of various methods in the second scenario provided in the embodiment of the present application. Fig. 9 This is a comparison chart of the tracking performance of various methods in the second scenario provided in the embodiment of the present application. Figures 6 to 9 In this paper, the truth method, TrMTT, DeepMTT, ISPM, HGMM and MBMM are used as examples. Figure 7 (a) Figure 7 (b) Figure 7 (c) and Figure 7 (d) Sharing Figure 7 The legend on the left uses different colored lines to represent the true value method, this method, TrMTT, DeepMTT, ISPM, HGMM, and MBMM. Fig. 9 (a) Fig. 9 (b) Fig. 9 (c) and Fig. 9 (d) Sharing Fig. 9 The legend on the right.

[0095] The truth value method is a trajectory tracking method that optimizes the difference between the true value and the estimated value. This method calculates the error between the actual trajectory of the target and the expected trajectory and continuously adjusts the controller parameters to achieve the purpose of accurate tracking. TrMTT is a high-maneuverable target radar tracking deep learning model based on the Transformer structure, while DeepMTT is a high-maneuverable target tracking algorithm based on deep learning of a bidirectional LSTM network, ISPM is an image signal processing algorithm, HGMM is a mixed Gaussian model, and MBMM is a detection-based multi-target tracking algorithm.

[0096] like Figure 6 and Figure 8 As shown, it includes the filtering effect diagram of each method described above on the image plane, and Figure 7 and Fig. 9They are respectively the error diagrams of the tracking filter sequence changing with time in the four dimensions of position and speed in the corresponding scenarios. On the trajectory after filtering, it can be seen that the ASDN of the present application can be closer to the real target trajectory (black). This is mainly due to the frequency domain modeling of the noise, which enables the target observation sequence to be denoised globally, avoiding the problem of traditional filtering algorithms obtaining cumulative errors due to noise and thus filtering divergence (the green trajectory is the effect diagram of the unscented Kalman filter). Not only that, when the local target makes a large maneuver turn, the local area is magnified and displayed. It can be seen that the present application can still track the target trajectory well, and achieve better performance than other methods in terms of position and speed estimation errors. This is mainly due to the targeted modeling of local features in the time domain by the present application, which can better perceive the nonlinear maneuvers that occur locally.

[0097] The high-mobility target tracking method based on frequency domain modeling and denoising network provided by this application has the smallest statistical error at all scales, and compared with traditional and current deep learning-based methods, the speed of this method is at least two orders of magnitude ahead of other algorithms. While achieving high performance, it also achieves extremely fast reasoning speed, which is mainly due to the overall architecture's comprehensive modeling of global frequency domain features and local time domain features. The high-mobility target tracking method based on frequency domain modeling and denoising network provided by this application can be deployed on a high-performance graphics processor to achieve end-to-end real-time output, greatly improving computing efficiency, and no human prior parameter adjustment is required.

[0098] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other. For the system provided in the embodiment, since it corresponds to the method provided in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0099] Specific examples are used herein to illustrate the principles and implementation methods of the present application, and the description of the above embodiments is only used to help understand the method and core ideas of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the present application.

[0100] It should also be noted that, in this specification, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

Claims

1. A high-mobility target tracking method based on frequency domain modeling and denoising network, characterized in that: include: Obtaining the observed trajectory sequence of highly maneuverable targets; Preprocessing the observation trajectory sequence to obtain a preprocessed sequence; The preprocessing sequence is input into an adaptive frequency domain noise reduction tracking network for tracking and filtering processing to obtain a frequency domain filtering result; wherein the adaptive frequency domain noise reduction tracking network is used to filter out noise, and extract the global features of the preprocessing sequence in the frequency domain, and extract the local features of the preprocessing sequence in the time domain; Tracking the high-maneuverability target according to the frequency domain filtering result; The step of inputting the preprocessing sequence into an adaptive frequency domain noise reduction tracking network for tracking filtering includes: In the adaptive frequency domain denoising tracking network, a timestamp component is added to each of the preprocessed sequences using an input layer, and the preprocessed sequences with the timestamp component added are mapped to a feature space through a linear layer; After the preprocessing sequence is mapped, it is input into the cascaded filter layer for processing to obtain a frequency domain filtering result; wherein the filter layer includes a frequency domain processing layer and a time domain processing layer; Among them, the processing of the filter layer of the input cascade includes: Transforming the preprocessed sequence into the frequency domain through a one-dimensional discrete Fourier transform; In the frequency domain, one of the variables is filtered by a global filter to obtain a first filtering result; the other variable is first filtered in the frequency domain by an adaptive mask containing a power spectrum map and then input into the global filter to obtain a second filtering result; The first filtering result and the second filtering result are superimposed to obtain the frequency domain filtering result.

2. The high-mobility target tracking method based on frequency domain modeling and denoising network according to claim 1 is characterized in that: Preprocessing the observation trajectory sequence includes: Performing coordinate system transformation on the observation trajectory sequence; Determining the velocity components of the high maneuverability target so that the input sequence feature dimension and the output dimension are consistent; and determining that the components of each dimension conform to a normal distribution; Normalization processing is performed on different physical quantities of the observation trajectory sequence.

3. The high-mobility target tracking method based on frequency domain modeling and denoising network according to claim 1 is characterized in that: The input is processed by the filter layer of the cascade, and the frequency domain filtering result is obtained, which also includes: Before being input into the time domain processing layer, the frequency domain filtering result is converted into the time domain through a one-dimensional inverse discrete Fourier transform.

4. The high-mobility target tracking method based on frequency domain modeling and denoising network according to claim 1 is characterized in that: The input is processed in the cascaded filter layer, and the frequency domain filtering results include: In the time domain processing layer, at least two convolutional layers with kernels of different sizes are used to extract time domain local features respectively; The two-way time domain local features are fused as the supervisory signals of the other way respectively, and the final features are obtained after fusion; The final feature is mapped back to the target real physical space to obtain a frequency domain filtering result.

5. The high-mobility target tracking method based on frequency domain modeling and denoising network according to claim 1 is characterized in that: Tracking the high maneuverability target according to the frequency domain filtering result includes: The frequency domain filtering result is generated to generate a filtering effect graph, and the high-maneuverability target is tracked according to the filtering effect graph.

6. The high-mobility target tracking method based on frequency domain modeling and denoising network according to claim 2 is characterized in that: The coordinate system conversion of the observation trajectory sequence includes: The observation trajectory sequence is converted from a polar coordinate system to a plane coordinate system according to the observation trajectory sequence, the converted time, the distance of the target from the sensor, and the observation angle.

7. The high-mobility target tracking method based on frequency domain modeling and denoising network according to claim 2 is characterized in that: Determining the velocity component of the high maneuverability target includes: The velocity component of the high maneuverable target is estimated using a time domain difference method; the time domain difference method is used to discretize the observed trajectory sequence in the time domain, use finite difference approximation to calculate the rate of change, and obtain an estimated value of the velocity component based on the rate of change.

8. The high-mobility target tracking method based on frequency domain modeling and denoising network according to claim 2 is characterized in that: Performing normalization processing on different physical quantities of the observation trajectory sequence includes: A normalization formula is used to perform normalization processing on different physical quantities of the observation trajectory sequence.

Images