Multi-frame joint detection and tracking method of complex background complex-valued u-shaped segmentation network

Abstract

The application discloses a complex background under a complex background, and discloses a complex U-shaped segmentation network multi-frame joint detection pre-tracking method. First, the collected radar data is processed to obtain RDA data cubes. Then, a complex U-shaped segmentation network is used to generate an accurate confidence cube, so as to distinguish targets from backgrounds. Then, based on the extracted target state, a more accurate state transition range is obtained, and a multi-frame pre-tracking algorithm is combined to determine the target path. The method of the application is aimed at radar RDA characteristic data, a multi-dimensional network feature extraction module is designed, the target segmentation effect of 3D data is significantly improved, and the complex U-shaped segmentation network is fully utilized to obtain more accurate target existence information, so that the problem that the targets cannot be effectively tracked due to the lack of clutter prior information under the complex background is solved. Based on the target confidence data, a confidence extraction and track state transition range are designed, the calculation complexity is reduced, and the problem of tracking loss caused by the mismatch of the motion model of the maneuvering target is solved.

Description

Technical Field

[0001] This invention belongs to the field of radar technology, specifically relating to a multi-frame joint detection and tracking method for complex-valued U-shaped segmentation networks under complex backgrounds. Background Technology

[0002] Compared to other sensors, radar possesses advantages such as all-weather operation, long-range detection capabilities, and high-precision measurement capabilities, making it particularly outstanding in detecting and tracking various targets in complex environments. Target detection and tracking technologies play a crucial role in civilian sensing systems within the radar field. By processing and analyzing radar data, the motion state, position, and number of targets in a scene can be estimated. Track-Before-Detect (TBD) technology, with its ability to detect low-observable targets in complex environments, has experienced rapid development and widespread application, becoming one of the current research hotspots in the radar field. By jointly processing multiple frames of raw echo data, it avoids single-frame detection misjudgments, effectively solving the problem of missed target detection, and features strong noise resistance and excellent tracking performance. Because multi-frame TBD can effectively detect and track targets under low signal-to-noise ratio conditions, it has broad development prospects as a target tracking technology.

[0003] For multi-frame pre-detection tracking techniques, a wider range of target path searches is achieved by increasing the number of samples in the time dimension. Although existing multi-frame pre-detection tracking techniques have the ability to detect low-observable targets in complex environments, they still have fundamental shortcomings. The accumulated statistics of existing multi-frame pre-detection tracking techniques are usually in the form of log-likelihood ratios, which requires assumptions about background noise in the environment to establish target and measurement models, and to set the basic parameters of these models. However, the background in real-world environments is complex, making it difficult to obtain background-related prior knowledge. Once the assumed model mismatches with the real scene, it will lead to unreasonable settings of the accumulated statistics, severely reducing the detection and tracking performance of the algorithm. In addition, the algorithm performance depends on the target motion model. When the motion model mismatches, the state transition range will no longer be accurate, thus failing to effectively search for the target state, resulting in a decline in detection and tracking performance.

[0004] The paper "Track-Before-Detect Strategies for Radar Detection in G0-Distributed Clutter, IEEE Trans Aero and Elect Sys, AES-53, 2516-2533, Oct. 2017" improves the multi-frame pre-detection tracking framework in non-Gaussian clutter environments, but this paper requires the clutter model to adopt a generalized G0 distribution. The paper "Complex Likelihood Ratio Pre-Detection Tracking Algorithm Based on Dynamic Programming in Non-Uniform Clutter Environments, Fire Control Radar Technology, 2020, 49(1): 49~55" proposes to re-derive the single-frame accumulation statistics based on the composite likelihood ratio function, but this scheme requires the combination of prior clutter map information. None of the above existing methods have fundamentally overcome the limitations of the TBD algorithm. They simply redesign the accumulation statistics and state transition range of TBD based on a wider range of models or conditions. The algorithm still requires prior clutter information, which is unavailable in real-world environments. The paper “RADGAN: Applying Adversarial Machine Learning to Track-Before-Detect Radar, 2021 IEEE RadarConference, 2021, pp. 1-6” uses Generative Adversarial Network (GAN) to perform binary classification on the data in radar images, thereby highlighting the region in the range Doppler image; the paper “Combining denoising autoencoders and dynamic programming for acoustic detection and tracking of underwater moving targets. Sensors, 2020, 20(10): 29-45” regards the deep convolutional denoising autoencoder (CDA) as a denoising filter and performs accumulation operation on the echo data after CDA processing, thereby realizing the detection and tracking of underwater targets; the paper “Dynamic Programming Track-Before-Detection Algorithm Based on Neural Network, Modern Radar, 2017, 39(11): 34~38” uses a neural network based on Hebb rules to perform cluster analysis on the original measurement data, thereby scaling the amplitude of the measurement data. The aforementioned deep learning processing methods essentially only pre-screen the amplitude features of the data, without performing multi-level mining of radar data, and cannot effectively solve the tracking anomaly problem when the noise model and motion model are mismatched. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a multi-frame joint detection and pre-tracking method using a complex-valued U-shaped segmentation network in complex backgrounds. This method solves the problem that existing methods fail to achieve target tracking in complex backgrounds due to the difficulty in obtaining prior information about clutter and the target, leading to mismatches in the value function and the target motion model.

[0006] The technical solution adopted in this invention is: a multi-frame joint detection and tracking method for complex-valued U-shaped segmentation networks under complex backgrounds, the specific steps of which are as follows:

[0007] S1. Construct a complex-valued U-shaped segmentation network model;

[0008] The complex-valued U-shaped segmentation network model includes an encoder layer on the left and a decoder layer on the right; and the complex-valued U-shaped segmentation network model uses Dice_CELoss as the loss function during the network training phase.

[0009] S2. Generate training datasets and label sets for millimeter-wave radar under different complex noise background conditions. After 3D-FFT signal processing, the training dataset is used to obtain the RDA cube training set, which is input together with the label set into the complex value U-shaped segmentation network model for network training to obtain the optimal weights of the model.

[0010] S3. After performing 3D-FFT signal processing on the original radar data received by the actual radar equipment frame by frame, the corresponding RDA cube is obtained. The cube is then input into the complex-valued U-shaped segmentation network and the optimal weights obtained in step S2 are used for prediction to obtain the radar measurement confidence cube.

[0011] S4. Based on the radar measurement confidence cube obtained in step S3 Confidence level extraction is performed to achieve the transformation from three-dimensional confidence level to two-dimensional XY confidence level plane, and the state transition range of target point trace is determined.

[0012] S5. Based on the MF-TBD algorithm framework, multiple frames of XY confidence plane data are jointly processed. Finally, the target trajectory is obtained through threshold decision and trajectory backtracking, realizing joint detection and tracking of multiple frames of data.

[0013] Furthermore, in step S1, the design of each module of the complex-valued U-shaped segmentation network model is as follows:

[0014] (1) Encoder layer;

[0015] Each stage of the encoder layer consists of repeated combinations of the feature extraction module UX-Net Block and the downsampling layer, as expressed below:

[0016] (1)

[0017] in, This indicates the feature extraction module. This indicates downsampling processing. , They represent the first Layer input and output features, and .

[0018] feature , , , and The dimensions are respectively , , , and The resolution gradually decreases from top to bottom, and the number of channels doubles in turn.

[0019] in, Representing RDA cubes respectively The length of the dimension.

[0020] (2) Feature extraction module;

[0021] The feature extraction module comprises two large-kernel convolutional blocks. The first large-kernel convolutional block includes: a large-kernel convolution (DWC), a depthwise convolutional scaling (DCS), residual connections, and a Gaussian error linear unit (GELU). The second large-kernel convolutional block includes: a large-kernel convolution (DWC), a depthwise convolutional scaling (DCS), and residual connections. The relationship between the two large-kernel convolutional blocks is expressed as follows:

[0022] (2)

[0023] in, and They represent the first Large kernel convolution and the first layer The output of the large kernel convolution block of the layer, Presentation layer normalization processing, This represents the GELU activation function, large kernel convolution. The size is Depth convolution scaling The kernel size is And input and output The size is the size of the corresponding coding layer feature. In the feature extraction module, the feature dimension size is not changed.

[0024] (3) Decoder layer;

[0025] The multi-scale output of each stage in the decoder is connected to the decoder layer through residual blocks and fused with the feature upsampling results of the previous stage. The number of stages corresponds to the number of encoder layers, and the process is expressed as follows:

[0026] (3)

[0027] in, This indicates residual processing, including two normalized values. Convolutional layer Indicates upsampling, Indicates the first Layer encoder output features Indicates the first Layer decoder output features , , , and The dimensions are respectively , , , and The resolution gradually increases from bottom to top, while the number of channels is halved in each subsequent step.

[0028] (4) Residual module;

[0029] The residual module includes two normalized... The specific expression for each convolutional layer is as follows:

[0030] (4)

[0031] in, Indicates the first Layer Convolutional layer weight tensor This represents the convolution operation. Indicates the first Layer bias terms. , These represent the input and output features of the convolutional layer, respectively. .

[0032] Furthermore, step S2 is specifically as follows:

[0033] Training datasets and label sets for millimeter-wave radar are generated under different complex noise background conditions, and the system parameters are initialized. Then, the training dataset of millimeter-wave radar is processed by 3D-FFT to obtain the RDA cube training set.

[0034] The expression for 3D-FFT signal processing is as follows:

[0035] (5)

[0036] (6)

[0037] (7)

[0038] in, Indicates the number of time sampling points. Indicates the number of chirps. Indicates the number of antenna channels; These represent the target's distance, Doppler shift, and angle, respectively. Indicates the first The nth time sample, the nth The antenna receiving channel, the first Radar data from one chirp. Represents the distance dimension FFT, Indicates distance-Doppler FFT, This represents a distance-Doppler-angle FFT. That is, the 3D-FFT result, RDA cube data.

[0039] Then, the RDA cube training set data is split into real-valued cubes. and imaginary cube Two channels are input into a complex-valued U-shaped segmentation network along with the label set to train the model and obtain the optimal weights.

[0040] Furthermore, step S3 is specifically as follows:

[0041] The raw millimeter-wave radar data is processed using 3D-FFT to obtain RDA cube data, which is then split into real-valued cubes. and imaginary cube The optimal weights of the model obtained in step S2 are used to adjust the received radar real-valued cube. Imaginary cube Prediction is performed, and the radar measurement confidence cube is finally obtained. .

[0042] The expression for the RDA cube data splitting process is as follows:

[0043] (8)

[0044] in, , and These represent the target's distance, Doppler shift, and angle, respectively. It represents the imaginary unit.

[0045] Furthermore, step S4 is specifically as follows:

[0046] For the confidence cube of the k-th frame, the corresponding target points are obtained through low-threshold detection. Reconstruct the XY plane of the corresponding frame and save the radial velocity information of the target point. The process expression is as follows:

[0047] (9)

[0048] in, Represents the reconstructed XY plane. and Indicates the first Frame point The corresponding XY coordinates, This represents the confidence threshold value. Represents a very small value. This represents the distance and angle coordinates on the confidence cube in the k-th frame. The expression for the transition process of the target point's state is as follows:

[0049] (10)

[0050] By the tangential velocity Perform interval traversal and combine it with the saved radial velocity information of the target point trace. The state transition range of the corresponding target point is obtained. The process expression is as follows:

[0051] (11)

[0052] (12)

[0053] in, Indicates the first The state of the target point at any given time. Indicates the measurement time interval.

[0054] Furthermore, step S5 is specifically as follows:

[0055] Based on the MF-TBD algorithm framework, multi-frame XY confidence plane data and corresponding target point state transition ranges are jointly processed to obtain the multi-frame cumulative value function plane. The process expression is as follows:

[0056] (13)

[0057] in, Represents the stage value function, Used to store inter-frame state transitions. Representing state The corresponding value on the XY plane.

[0058] Finally, the target trajectory is obtained through threshold decision and trajectory backtracking. The process expression is as follows:

[0059] (14)

[0060] (15)

[0061] (16)

[0062] in, Indicates the threshold value. Indicates the first The tracking target state of the frame.

[0063] The beneficial effects of this invention are as follows: The method of this invention first processes the acquired radar data to obtain an RDA data cube, then uses a complex-valued U-shaped segmentation network to generate an accurate confidence cube, thereby distinguishing the target from the background. Based on the extracted target state, a more accurate state transition range is obtained, and the target path is jointly determined by combining a multi-frame detection preceding aggregator algorithm. This invention's method, targeting radar RDA feature data, designs a multi-dimensional network feature extraction module, significantly improving the target segmentation effect of 3D data. It also fully utilizes the complex-valued U-shaped segmentation network to obtain more accurate target presence information, solving the problem of ineffective tracking caused by the lack of prior information from clutter in complex backgrounds. Furthermore, based on target confidence data, it designs confidence extraction and point state transition ranges, reducing computational complexity while solving the tracking loss problem caused by the mismatch of the motion model of the maneuvering target.

[0064] The method of this invention overcomes the problems of value function mismatch caused by insufficient prior information in complex backgrounds and tracking failure when the target maneuvers exceed the state transition range in existing traditional pre-detection tracking algorithms. Thus, it achieves accurate identification and tracking of small maneuvering targets in complex backgrounds and can be applied to maneuvering target tracking in complex dynamic backgrounds such as oceans, atmospheres and cities. Attached Figure Description

[0065] Figure 1 This is a flowchart of a multi-frame joint detection and tracking method for complex-valued U-shaped segmentation networks under complex backgrounds, according to the present invention.

[0066] Figure 2 This is a structural diagram of the complex-valued U-shaped segmentation network model in an embodiment of the present invention.

[0067] Figure 3 This is an architecture diagram of the feature extraction module in the complex-valued U-shaped segmentation network model of this invention.

[0068] Figure 4 This is a flowchart of 3D confidence extraction and tracking in an embodiment of the present invention.

[0069] Figure 5 This is a schematic diagram of the XY plane background under the complex background and radar parameter configuration in Table 1 in an embodiment of the present invention.

[0070] Figure 6 This is a schematic diagram of the XY plane background under the complex background and radar parameter configuration in Table 2 in an embodiment of the present invention.

[0071] Figure 7 This is a schematic diagram of the XY plane motion of the maneuvering target under a complex background as shown in Table 1 in an embodiment of the present invention.

[0072] Figure 8 This is a schematic diagram of the XY plane motion of the maneuvering target under a complex background as shown in Table 2 in an embodiment of the present invention.

[0073] Figure 9 This is a schematic diagram of the multi-frame accumulated value function of a maneuvering target under the complex background and radar parameter configuration in Table 1 in an embodiment of the present invention.

[0074] Figure 10 This is a schematic diagram of the multi-frame accumulated value function of a maneuvering target under the complex background and radar parameter configuration in Table 2 in an embodiment of the present invention.

[0075] Figure 11 This is a comparison chart of the tracking accuracy of the method of the present invention and the existing method under complex background and radar parameter configurations in Table 1 for 300 trials in an embodiment of the present invention.

[0076] Figure 12 This is a comparison chart of the tracking accuracy of the method of the present invention and the existing method under complex background and radar parameter configurations in Table 2 for 300 trials in an embodiment of the present invention. Detailed Implementation

[0077] To facilitate the description of the method of the present invention, the following terms will first be explained:

[0078] Term 1: Radar raw data;

[0079] Raw radar data refers to unprocessed signal data directly collected by radar sensors.

[0080] Term 2: RDA cube;

[0081] An RDA cube is a data cube that has been processed from raw data. It includes information in three dimensions: range, azimuth, and Doppler shift, and is used to characterize the frequency domain features of radar signals.

[0082] Term 3: Confidence level;

[0083] Radar data, after processing, is used to characterize the probability of a target's presence. A higher confidence value indicates a greater likelihood of the target's existence.

[0084] Term 4: Complex background;

[0085] Complex backgrounds refer to fluctuating background noise, arbitrary background noise, etc., where prior information is unknown.

[0086] Term 5: Radial velocity;

[0087] Radial velocity is the velocity component of the target relative to the radar, along the radar beam direction.

[0088] Term 6: Tangential velocity;

[0089] Tangential velocity is the velocity component of the target perpendicular to the radial velocity.

[0090] Term 7: XY plane;

[0091] The XY plane refers to a two-dimensional plane representation obtained in the Cartesian coordinate system by transforming the distance and angle information of the target point.

[0092] The method of this invention is mainly verified through simulation experiments. All steps and conclusions have been verified to be correct using Matlab 2020 and PyCharm 2023. The method of this invention will be further described below with reference to the accompanying drawings and embodiments.

[0093] like Figure 1 The flowchart shown is a multi-frame joint detection and tracking method for complex-valued U-shaped segmentation networks under complex backgrounds according to the present invention. The specific steps are as follows:

[0094] S1. Construct a complex-valued U-shaped segmentation network model;

[0095] like Figure 2 As shown, the complex-valued U-shaped segmentation network model includes an encoder layer on the left and a decoder layer on the right; and the complex-valued U-shaped segmentation network model uses Dice_CELoss as the loss function during the network training phase.

[0096] S2. Generate training datasets and label sets for millimeter-wave radar under different complex noise background conditions. After 3D-FFT signal processing, the training dataset is used to obtain the RDA cube training set, which is input together with the label set into the complex value U-shaped segmentation network model for network training to obtain the optimal weights of the model.

[0097] S3. After performing 3D-FFT signal processing on the original radar data received by the actual radar equipment frame by frame, the corresponding RDA cube is obtained. The cube is then input into the complex-valued U-shaped segmentation network and the optimal weights obtained in step S2 are used for prediction to obtain the radar measurement confidence cube.

[0098] S4. Based on the radar measurement confidence cube obtained in step S3 Confidence level extraction is performed to achieve the transformation from three-dimensional confidence level to two-dimensional XY confidence level plane, and the state transition range of target point trace is determined.

[0099] S5. Based on the MF-TBD algorithm framework, multiple frames of XY confidence plane data are jointly processed. Finally, the target trajectory is obtained through threshold decision and trajectory backtracking, realizing joint detection and tracking of multiple frames of data.

[0100] like Figure 2 As shown, in this embodiment, the design of each module of the complex-valued U-shaped segmentation network model in step S1 is as follows:

[0101] (1) Encoder layer;

[0102] Each stage of the encoder layer consists of repeated combinations of a feature extraction module (UX-Net Block) and a downsampling layer, and the process is expressed as follows:

[0103] (1)

[0104] in, This indicates the feature extraction module. This indicates downsampling processing. , They represent the first Layer input and output features, and .

[0105] feature , , , and The dimensions are respectively , , , and The resolution gradually decreases from top to bottom, and the number of channels doubles in turn.

[0106] in, Representing RDA cubes respectively The length of the dimension. In this embodiment... The sizes are 128, 64, and 128 respectively.

[0107] (2) Feature extraction module;

[0108] like Figure 3 As shown, the feature extraction module comprises two large-kernel convolutional blocks. The first large-kernel convolutional block within the module includes: large-kernel convolution (DWC), depthwise convolution scaling (DCS), residual connections, and Gaussian error linear units (GELUs). The second large-kernel convolutional block includes: large-kernel convolution (DWC), depthwise convolution scaling (DCS), and residual connections. The relationship between the two large-kernel convolutional blocks is expressed as follows:

[0109] (2)

[0110] in, and They represent the first Large kernel convolution and the first layer The output of the large kernel convolution block of the layer, Presentation layer normalization processing, This represents the GELU activation function, large kernel convolution. The size is Depth convolution scaling The kernel size is And input and output The size is the size of the corresponding coding layer feature. In the feature extraction module, the feature dimension size is not changed.

[0111] (3) Decoder layer;

[0112] The multi-scale output of each stage in the decoder is connected to the decoder layer through residual blocks and fused with the feature upsampling results of the previous stage. The number of stages corresponds to the number of encoder layers, and the process is expressed as follows:

[0113] (3)

[0114] in, This indicates residual processing, including two normalized values. Convolutional layer Indicates upsampling, Indicates the first Layer encoder output features Indicates the first Layer decoder output features , , , and The dimensions are respectively , , , and The resolution gradually increases from bottom to top, while the number of channels is halved in each subsequent step.

[0115] (4) Residual module;

[0116] The residual module includes two normalized... The specific expression for each convolutional layer is as follows:

[0117] (4)

[0118] in, Indicates the first Layer Convolutional layer weight tensor This represents the convolution operation. Indicates the first Layer bias terms. , These represent the input and output features of the convolutional layer, respectively. .

[0119] In this embodiment, step S2 is specifically as follows:

[0120] Training datasets and label sets for millimeter-wave radar are generated under different complex noise background conditions, and the parameters of each system are initialized. Then, the training dataset of millimeter-wave radar is processed by 3D-FFT (three-dimensional Fourier transform) to obtain the RDA cube training set.

[0121] The expression for 3D-FFT signal processing is as follows:

[0122] (5)

[0123] (6)

[0124] (7)

[0125] in, Indicates the number of time sampling points. Indicates the number of chirps. Indicates the number of antenna channels; These represent the target's distance, Doppler shift, and angle, respectively. Indicates the first The nth time sample, the nth The antenna receiving channel, the first Radar data from one chirp. Represents the distance dimension FFT, Indicates distance-Doppler FFT, This represents a distance-Doppler-angle FFT. That is, the 3D-FFT result, RDA cube data.

[0126] Then, the RDA cube training set data is split into real-valued cubes. and imaginary cube Two channels are input into a complex-valued U-shaped segmentation network along with the label set to train the model and obtain the optimal weights.

[0127] In this embodiment, step S3 is specifically as follows:

[0128] The raw millimeter-wave radar data is processed using 3D-FFT (the 3D-FFT signal processing process is the same as step S2) to obtain RDA cube data, which is then split into real-valued cubes. and imaginary cube The optimal weights of the model obtained in step S2 are used to adjust the received radar real-valued cube. Imaginary cube Prediction is performed, and the radar measurement confidence cube is finally obtained. .

[0129] The expression for the RDA cube data splitting process is as follows:

[0130] (8)

[0131] in, , and These represent the target's distance, Doppler shift, and angle, respectively. It represents the imaginary unit.

[0132] like Figure 4 As shown, in this embodiment, step S4 is specifically as follows:

[0133] For the confidence cube of the k-th frame, the corresponding target points are obtained through low-threshold detection. Reconstruct the XY plane of the corresponding frame and save the radial velocity information of the target point. The process expression is as follows:

[0134] (9)

[0135] in, Represents the reconstructed XY plane. and Indicates the first Frame point The corresponding XY coordinates, This represents the confidence threshold value. Represents a very small value. This represents the distance and angle coordinates on the confidence cube in the k-th frame. The expression for the transition process of the target point's state is as follows:

[0136] (10)

[0137] By the tangential velocity Perform interval traversal and combine it with the saved radial velocity information of the target point trace. The state transition range of the corresponding target point is obtained. The process expression is as follows:

[0138] (11)

[0139] (12)

[0140] in, Indicates the first The state of the target point at any given time. Indicates the measurement time interval.

[0141] like Figure 4 As shown, in this embodiment, step S5 is specifically as follows:

[0142] Based on the MF-TBD algorithm framework, multi-frame XY confidence plane data and corresponding target point state transition ranges are jointly processed to obtain the multi-frame cumulative value function plane. The process expression is as follows:

[0143] (13)

[0144] in, Represents the stage value function, Used to store inter-frame state transitions. Representing state The corresponding value on the XY plane.

[0145] Finally, the target trajectory is obtained through threshold decision and trajectory backtracking. The process expression is as follows:

[0146] (14)

[0147] (15)

[0148] (16)

[0149] in, Indicates the threshold value. Indicates the first The tracking target state of the frame.

[0150] This embodiment also included experimental verification. Tables 1 and 2 are the radar and algorithm parameter configuration tables for complex background fluctuating noise environments and non-uniform clutter environments, respectively. The corresponding noise diagrams in the XY plane are shown below. Figure 5 , Figure 6 The trajectory of the moving target is as follows Figure 7 , Figure 8 .

[0151] Table 1

[0152]

[0153] in, Indicates wavelength.

[0154] Table 2

[0155]

[0156] Finally, multi-frame joint detection and target tracking under complex backgrounds were completed. The K-frame confidence plane value function accumulation results of the method of this invention are as follows: Figure 9 , Figure 10 As shown, Figure 11 , Figure 12 This figure shows the accuracy curve of the target detection in 300 Monte Carlo simulations of the method of the present invention in this embodiment. "Proposed" represents the target detection accuracy of the method of the present invention, and "DP-TBD" represents the target detection accuracy of existing traditional pre-detection tracking. As can be seen from this embodiment, the method of the present invention can effectively achieve multi-frame joint detection and tracking of targets in complex backgrounds.

[0157] In summary, the method of this invention achieves in-depth feature mining by transforming the raw radar data at multiple scales and designing a complex-valued U-shaped segmentation network to obtain accurate single-frame target confidence, thereby solving the problem of ineffective target tracking caused by insufficient prior information and statistical mismatch. Simultaneously, it proposes a method for constructing state transition ranges based on point state information, obtaining more accurate predicted points and overcoming tracking failures caused by target maneuverability. This invention utilizes a single-time-dimensional confidence cube obtained by the segmentation network to replace the traditional value function and determines a new state transition range based on point information. Combined with a multi-frame joint detection algorithm, it ultimately achieves target detection and trajectory backtracking, significantly improving the tracking accuracy of weak, maneuvering targets in complex backgrounds. It is applicable to maneuvering target tracking in complex dynamic backgrounds such as oceans, atmospheres, and cities.

[0158] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles and effects of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the essence of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A complex background under complex-valued U-shaped segmentation network multi-frame joint detection front tracking method, the specific steps are as follows: S1, constructing a complex-valued U-shaped segmentation network model; The complex U-shaped segmentation network model comprises: The left encoder layer and the right decoder layer; and the complex-valued U-shaped segmentation network model adopts Dice_CELoss as the loss function in the network training stage; S2, generate the training data set and the label set of the millimeter wave radar under different complex noise background conditions, and obtain the RDA cube training set after 3D-FFT signal processing of the training data set, input the complex-valued U-shaped segmentation network model together with the label set to obtain the optimal weight value of the model through network training; S3, sequentially frame by frame 3D-FFT signal processing is performed on the radar raw data received by the actual radar equipment to obtain the corresponding RDA cube, input the complex-valued U-shaped segmentation network, and use the optimal weight value obtained in step S2 to predict to obtain the radar measurement confidence cube; S4, based on the radar measurement confidence cube obtained in step S3 confidence extraction is performed to realize conversion from three-dimensional confidence to two-dimensional X-Y confidence plane and determine target track state transition range; S5, based on the MF-TBD algorithm framework, jointly process multi-frame X-Y confidence plane data, and finally obtain the target track through threshold decision and track backtracking, realize the joint detection and tracking of multi-frame data.

2. The complex background complex-valued U-shaped segmentation network multi-frame joint detection pre-tracking method according to claim 1, characterized in that, In step S1, the modules of the complex-valued U-shaped segmentation network model are designed as follows: (1) encoder layer; Each stage of the encoder layer is composed of a feature extraction module UX-Net Block and a downsampling layer, and the process expression is as follows: （1）； wherein, denotes a feature extraction module, denotes a down-sampling process, , denote the input and output features of the first layer, respectively, and ; Features , , , and The size of , , , and , from top to bottom, the resolution gradually decreases, and the number of channels doubles in turn; wherein, RDA cube length of a dimension; (2) feature extraction module; The feature extraction module includes two large kernel convolution blocks, the first layer of which includes: large kernel convolution DWC, depth convolution scaling DCS, residual connection and Gaussian error linear unit GELU, and the second layer of which includes: large kernel convolution DWC, depth convolution scaling DCS, residual connection; the relationship between the two large kernel convolution blocks is expressed as follows: （2）； in, and They represent the first Large kernel convolution and the first layer The output of the large kernel convolution block of the layer, Presentation layer normalization processing, This represents the GELU activation function, large kernel convolution. The size is Depth convolution scaling The kernel size is And input and output The size is the size of the corresponding coding layer feature; in the feature extraction module, the feature dimension size is not changed. (3) decoder layer; The multi-scale output of each stage in the decoder is connected to the decoder layer through a residual block, and is fused with the feature up-sampling result of the previous stage, and the stage number corresponds to the encoder layer, and the process expression is as follows: （3）； wherein, denotes residual processing, including two normalized convolutional layers, denotes up-sampling, denotes the first layer encoder output feature, denotes the first layer decoder output feature, , , , and have sizes , , , and , respectively, with resolution gradually increasing from bottom to top and channel number halved successively; (4) residual module; The residual module includes two normalized convolutional layers, each layer being specifically expressed as follows: （4）； wherein, denotes the layer's convolutional layer weight tensor, denotes a convolution operation, denotes the layer's bias term; , denote the convolutional layer's input and output features, respectively, and .

3. The complex background complex-valued U-shaped segmentation network multi-frame joint detection pre-tracking method according to claim 1, characterized in that, Step S2 is as follows: Under different complex noise background conditions, generate the training data set and the label set of the millimeter wave radar, and initialize the system parameters, then process the training data set of the millimeter wave radar through 3D-FFT to obtain the RDA cube training set; The expression of 3D-FFT signal processing is as follows: （5）； （6）； （7）； wherein, denotes the number of time samples, denotes the number of chirps, denotes the number of antenna channels; denotes the range, Doppler shift, angle of the target, respectively; denotes the radar data of the th time sample, the th antenna receive channel, the th chirp; denotes the range dimension FFT, denotes the range-Doppler dimension FFT, denotes the range-Doppler-angle dimension FFT, i.e. 3D-FFT result, RDA cube data; Then the RDA cube training set data is split into real-value cube and imaginary-value cube Two channels are input into the complex U-shaped segmentation network together with the label set, and the optimal weight of the model is trained.

4. The complex background complex-valued U-shaped segmentation network multi-frame joint detection pre-tracking method according to claim 1, characterized in that, Step S3 is as follows: The raw millimeter-wave radar data is processed using 3D-FFT to obtain RDA cube data, which is then split into real-valued cubes. and imaginary cube The optimal weights of the model obtained in step S2 are used to adjust the received radar real-valued cube. Imaginary cube Prediction is performed, and the radar measurement confidence cube is finally obtained. ; Wherein, the RDA cube data splitting process expression is as follows: （8）； wherein, , and denote the range, Doppler shift and angle of the target, respectively, denotes the imaginary unit.

5. The complex background complex-valued U-shaped segmentation network multi-frame joint detection pre-tracking method according to claim 1, characterized in that, Step S4 is as follows: For the kth frame confidence cube, the corresponding target track is obtained by low threshold detection , the X-Y plane of the corresponding frame is reconstructed, and the radial velocity information of the target track is saved , the process expression is as follows: （9）； wherein, denotes the reconstructed X-Y plane, and denotes the k-th frame, frame point corresponding X-Y coordinates, denotes the confidence threshold value, denotes the extremely small value, denotes the distance, angle coordinates on the k-th frame confidence cube; the conversion process expression of the target point trace state is as follows: （10）； By traversing the interval of tangential velocity and combining the saved radial velocity information of the target track , the state transition range of the corresponding target track is obtained The process expression is as follows: （11）； （12）； wherein, represents the target track state at the time instant, represents the measurement time interval.

6. The complex background complex-valued U-shaped segmentation network multi-frame joint detection pre-tracking method according to claim 1, characterized in that, Step S5 is as follows: Based on the MF-TBD algorithm framework, jointly process multi-frame X-Y confidence plane data and the corresponding target point track state transition range to obtain a multi-frame accumulation value function plane, and the process expression is as follows: （13）； wherein, represents a stage value function, for storing inter-frame state transitions, represents a state values on the corresponding X-Y plane; Finally, the target track is obtained by threshold decision and track back The process expression is as follows: （14）； （15）； （16）； wherein, denotes a threshold value, denotes a first frame of the tracking target state.

Images