A multi-target detection pre-tracking method based on transunet segmentation guidance
By combining the TransUnet network and Huber loss function with dynamic programming algorithm, the problems of high computational complexity and high false alarm rate of traditional detection and tracking algorithms in low signal-to-noise ratio environments are solved, and efficient tracking of weak targets is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
- Filing Date
- 2026-03-18
- Publication Date
- 2026-07-03
AI Technical Summary
Traditional detection and tracking algorithms suffer from high computational complexity, discontinuous intelligent segmentation trajectories, and high false alarm rates in low signal-to-noise ratio environments, making it difficult to effectively track weak targets.
The TransUnet network is used to extract potential target candidate regions, the Huber loss function is used to fit the trajectory features, and a dynamic programming algorithm is used to make threshold decisions in the local state space. Fine trajectory estimation is performed through kinematic constraints.
It significantly improves the performance and robustness of weak target detection in low signal-to-noise ratio environments, reduces computational complexity, reduces false alarm interference, and enhances detection sensitivity and real-time performance.
Smart Images
Figure CN122330864A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radar target tracking, specifically relating to a multi-target detection pre-tracking method based on TransUnet segmentation guidance. Background Technology
[0002] Radar is an electronic device that uses emitted electromagnetic waves and received reflected signals to detect and locate targets, enabling long-range target detection around the clock and in all weather conditions. Traditional detection-tracking algorithms typically consist of two stages: detection and tracking. The detection stage performs threshold decisions on each frame of echo data, forming point data; the tracking stage performs correlation, filtering, and track management on the point data that exceeds the threshold, ultimately estimating the target's trajectory and achieving target tracking. Therefore, traditional detection-tracking algorithms are also known as Detect-Before-Track (DBT). Because it uses single-frame threshold detection, under the condition of ensuring a certain false alarm rate, it can lead to missed detections of targets with low signal-to-noise ratios and low signal-to-clutter ratios.
[0003] To improve the detection and tracking performance of weak targets by radar, it is also necessary to enhance the radar's ability to detect and track weak targets from the perspective of signal processing. Track-Before-Detect (TBD) accumulates and processes multiple frames of radar data, and takes advantage of the differences between target echoes and noise clutter in a higher-dimensional signal space to fully extract target information from the echoes, effectively improving the detection performance of weak targets.
[0004] Dynamic programming was proposed by American mathematician Richard Bellman in his research on optimization problems in multi-stage decision-making processes. The Bellman equation is the core of this theory, transforming a multi-stage process into a series of single-stage problems to be solved one by one. Domestic and international scholars have researched detection-pre-tracking based on dynamic programming. Yi Wei proposed a new dynamic programming algorithm for multi-target detection-pre-tracking, which effectively solves problems such as the curse of dimensionality and interference between nearby targets in multi-target tracking. The TransUNet network was first proposed by Chen. In the detection-pre-tracking of weak targets, Wu Xingrui's team designed a tracking algorithm for maneuvering weak targets applicable to different motion states based on the TransUNet model, achieving the detection and trajectory backtracking of weak maneuvering targets. This method verifies that using a semantic segmentation network model can effectively detect and track weak maneuvering targets in various motion states, with lower computational complexity than traditional methods and wider application scenarios.
[0005] To address the excessive computational burden of traditional pre-detection tracking algorithms and the model uncertainty exhibited by image segmentation using intelligent algorithms, this invention proposes a multi-target pre-detection tracking method based on TransUNet segmentation guidance. This method extracts potential target candidate regions through the TransUNet network and combines this with a robust M-fit method based on the Huber loss function to extract track features. Finally, the kinematically constrained DP-TBD algorithm performs threshold decisions on all candidate targets in the original data, significantly improving the detection performance and robustness of weak targets in low signal-to-noise ratio environments. Summary of the Invention
[0006] Purpose of the invention: This invention proposes a multi-target detection pre-tracking method based on TransUnet segmentation guidance to overcome the shortcomings of existing technologies in weak multi-target tracking under low signal-to-noise ratio environments, such as high computational complexity, discontinuous intelligent segmentation trajectories, and high false alarm rate.
[0007] Technical solution: The present invention provides a multi-target detection pre-tracking method based on TransUnet segmentation guidance, comprising the following steps:
[0008] (1) Construct a motion model and use maximum value accumulation to perform non-coherent superposition of multi-frame measurement data;
[0009] (2) The TransUnet network is used to segment the measurement data plane, extract global context features, and generate candidate regions for potential target trajectories;
[0010] (3) Based on Huber loss, M-estimation fits the potential target trajectory and extracts the starting point, direction of motion, and velocity of the potential target trajectory in parallel;
[0011] (4) Using the extracted position, velocity and direction parameters as prior guidance, a local state space under kinematic constraints is constructed based on multi-frame measurement data; the trajectories of all candidate regions are estimated in detail using dynamic programming before detection tracking algorithm within the local state space.
[0012] (5) Perform trajectory backtracking and threshold detection on the estimation results of step (4).
[0013] Furthermore, the process of constructing the motion model in step (1) is as follows:
[0014] The target moving in space The motion state at any given time consists of motion parameters including the target's position and velocity, where K is the total number of frames processed in a batch by the tracking algorithm. Its state variables are:
[0015]
[0016] in, , The position of the target point in the X and Y dimensions. , These are the velocities of the target point in the X and Y directions, respectively.
[0017] Consider a moving target in the XY plane. Its motion state evolution follows a one-dimensional Markov process, and its motion model is expressed as:
[0018]
[0019] Where Q represents process noise. Representing the inter-frame interval, when the target is moving at a constant velocity in a straight line in the XY plane, the above formula is:
[0020]
[0021] Where F is the state transition matrix:
[0022]
[0023] in, It is a second-order identity matrix. Let the Kronecker product be represented; then the track of a target in a batch of K frames is represented as:
[0024]
[0025] For the motion model, each measurement unit is a two-dimensional pixel plane, assuming the size of this pixel plane is... Then the measurement plane at time k Represented as a The matrix:
[0026]
[0027] Use respectively and Let represent the scenarios where the target exists and the scenario where it does not exist, respectively. The statistical characteristics of each resolution unit are given by the following formula:
[0028]
[0029] in, It is independent and identically distributed zero-mean complex Gaussian white noise. For constant amplitude and phase in A complex random variable whose internal distribution follows a uniform distribution.
[0030] Furthermore, the process of non-coherently superimposing multi-frame measurement data using maximum value accumulation in step (1) is as follows:
[0031] The measurement data plane is represented as follows:
[0032]
[0033] In the pixel unit Represented as:
[0034] .
[0035] Furthermore, the implementation process of step (2) is as follows:
[0036] The preprocessed measurement data is input into the TransUnet network, passing through an encoder, skip connections, and a decoder in sequence to obtain the intelligent segmentation result of the measurement plane. In the encoder stage, the accumulated echo data is first downsampled stepwise through three layers of convolution and pooling operations. Then, the obtained feature map is subjected to block embedding and position encoding to extract target features and perform pixel-level classification of the target and background based on the features. Further, the embedded one-dimensional vector is input into the Transformer module for hybrid encoding. Finally, through skip connections, the encoded feature map is sent to the corresponding decoder layer and concatenated with the data recovered by deconvolution upsampling to recover the corresponding pixel information in the measurement data plane.
[0037] During the nth layer processing of the encoder, given the input features Its output is The intermediate process value is denoted as The calculation of this layer consists of two core steps, expressed as follows:
[0038]
[0039]
[0040] Among them, MSA stands for Multi-head Self-Attention Module, MLP stands for Multilayer Perceptron Module, and LN represents the normalization operation of the layer.
[0041] Furthermore, the implementation process of the multi-head self-attention module (MSA) is as follows:
[0042] If the normalization function is softmax, then the output vector at the i-th position in the sequence is represented as:
[0043]
[0044] The query Q, key K, and value V matrices are formed by combining the input features X with learnable weight matrices. , , Obtained by linear transformation, i.e. , , ; It is the dimension of vector K; by calculating the correlation between Q and each K, the weight coefficient corresponding to each V is obtained, and then the V are weighted and summed to obtain the final attention value;
[0045] Multi-head self-attention learns multiple independent Q, K, V matrices in parallel, and calculates self-attention for each:
[0046]
[0047] Here, `head` is a self-attention head. The outputs of all `head`s are concatenated along the channel dimension and fused using a linear projection matrix `W`.
[0048]
[0049] Where h is the number of self-attention heads in multi-head self-attention.
[0050] Furthermore, the implementation process of step (3) is as follows:
[0051] Suppose that in the output binary mask, the k-th candidate connected region contains N pixels, and this set of pixels is represented as... Each pixel This represents its position in the cumulative image coordinate system;
[0052] Introducing the M-estimation theory, a regression model is constructed using the Huber loss function, which is insensitive to outliers: a fitted straight line L is constructed, which is formed by a point on the line... and unit direction vector The goal of optimization is to find the optimal parameters. This minimizes the sum of Huber losses of the residuals of all pixels to the line;
[0053] objective function The definition is as follows:
[0054]
[0055] in, For the i-th pixel The fitting residual to line L:
[0056]
[0057] The Huber loss function is defined as a piecewise function:
[0058]
[0059] When the residual When, the loss function is expressed as The norm ensures the fitting accuracy for interior points, when the residuals At that time, the loss function is expressed as Norms limit the influence of edge outlier noise, where This is a preset threshold parameter;
[0060] The objective function is solved by iterative reweighted least squares method, thus obtaining the direction of motion. and center point C;
[0061] Calculate each pixel in the point set P In the direction of the main axis of motion scalar projection value on :
[0062]
[0063] Iterate through all projected values and extract the minimum value. With the maximum value And based on this, reconstruct the two geometric endpoints of the trajectory. and :
[0064] .
[0065] Furthermore, the implementation process of step (4) is as follows:
[0066] Establish a local search space: Let the original measurement data be... For the k-th frame, the center position of the candidate target is predicted based on the uniform motion model. :
[0067]
[0068] Define the local search region for the k-th frame. For A rectangular area centered at R with a search radius of R:
[0069]
[0070] Where R is the preset search radius, used to deal with possible errors in parameter extraction during the feature extraction step;
[0071] when At that time, for all discrete states, we have:
[0072]
[0073]
[0074] when Then, the following calculations are performed on all discrete states:
[0075]
[0076]
[0077] in, Represents a value function. This represents the detection statistics for a single frame. Indicates the transition relationship between states. express The set of state transitions at time t, To smooth out the penalty coefficient, This represents the Euclidean distance between the current state and the state of the previous frame.
[0078] Furthermore, the implementation process of step (5) is as follows:
[0079] Threshold decision: when At that time, the following calculations are performed:
[0080]
[0081] in, The detection threshold is calculated from the false alarm rate and signal-to-noise ratio. If the conditions of the above formula are met, the trajectory is backtracked; if not, the trajectory reconstruction is declared to have failed.
[0082] Flight path retracing: Order Perform the following operations:
[0083]
[0084] Obtain the discrete trajectory of the target .
[0085] Beneficial Effects: Compared with the prior art, the beneficial effects of this invention are as follows: Compared with the traditional full-space traversal search (TBD) method, this invention uses DP-TBD to perform threshold decision on all candidate targets in the local state space, eliminating the interference of false tracks caused by network segmentation, ensuring the tracking performance of weak targets, reducing the computational complexity of weak target detection, and balancing real-time performance and sensitivity; at the same time, it effectively reduces the state search space by utilizing deep learning priors, significantly reducing computational overhead, effectively suppressing false alarm interference introduced by deep learning, and improving detection confidence. Attached Figure Description
[0086] Figure 1 This is a flowchart of the present invention;
[0087] Figure 2 Figures show the TransUNet network segmentation results under different signal-to-noise ratios; where (a) is the measurement space plane, (b) is the TransUNet network segmentation result, and (c) is the mask.
[0088] Figure 3 The track detection probabilities are compared between the intelligent segmentation guidance algorithm and the Hough transform method under different signal-to-noise ratios.
[0089] Figure 4 The algorithm complexity of the intelligent segmentation guidance algorithm is compared with that of the traditional DP-TBD algorithm at different frame numbers. Detailed Implementation
[0090] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0091] This invention proposes a multi-target detection pre-tracking method based on TransUNet segmentation guidance. Addressing the challenge of balancing computational complexity and detection performance in weak multi-target tracking under low signal-to-noise ratio environments, this invention proposes a method that utilizes network segmentation to obtain prior information about trajectory points and then reconstructs local spatial trajectories using Dynamic Programming-based Track-before-Detect (DP-TBD). In the measurement space generation stage, multi-frame measurement data are first non-coherently superimposed using maximum accumulation, and the accumulated image is input into the TransUNet network to extract global contextual features, generating candidate regions for potential target trajectories. To address the noise and track fragmentation issues inherent in network segmentation, a trajectory parameter extraction algorithm based on robust M-estimation is derived. The Huber loss function is used to suppress outliers and fit the principal axis of motion, and a bidirectional motion hypothesis is constructed to resolve the directional ambiguity caused by temporal loss. In the target tracking stage, the extracted position, velocity, and orientation parameters are used as prior guidance. Based on multi-frame measurement data, a local state space under kinematic constraints is constructed. All candidate trajectories are then used within this local space for small-scale, fine-grained track estimation using the Dynamic Programming-based Track-before-Detect (DP-TBD) algorithm. like Figure 1 As shown, the specific implementation process is as follows:
[0092] Step 1: Construct a motion model and use maximum accumulation to non-coherently stack multi-frame measurement data.
[0093] A target moving in space, its first... The motion state at any given time is composed of motion parameters such as the target's position and velocity, where K is the total number of frames processed in a batch by the tracking algorithm. This state variable is:
[0094]
[0095] in, , The position of the target point in the X and Y dimensions. , These are the velocities of the target point in the X and Y directions, respectively.
[0096] Consider a moving target in the XY plane. Its motion state evolution follows a one-dimensional Markov process, and its motion model can be expressed as:
[0097]
[0098] Where Q represents process noise. Representing the inter-frame interval, when the target is moving at a constant velocity in a straight line in the XY plane, the above formula can be written as:
[0099]
[0100] Where F is the state transition matrix:
[0101]
[0102] in, It is a second-order identity matrix, and This represents the Kronecker product. At this point, the trajectory of a target in a batch of K frames can be represented as:
[0103]
[0104] For this motion model, each measurement unit is a two-dimensional pixel plane, assuming the size of this pixel plane is... Then the measurement plane at time k It can be represented as a The matrix:
[0105]
[0106] Use respectively and Let represent the scenarios where the target exists and the scenario where it does not exist, respectively. The statistical characteristics of each resolution unit are given by the following formula:
[0107]
[0108] in, It is independent and identically distributed zero-mean complex Gaussian white noise. For constant amplitude and phase in A complex random variable whose internal distribution follows a uniform distribution.
[0109] The measurement data is preprocessed using the maximum value accumulation method. The measurement data plane can be represented as follows:
[0110]
[0111] In the pixel unit It can be represented as:
[0112] .
[0113] Step 2: Planar segmentation and feature extraction of measurement data based on TransUnet network.
[0114] The preprocessed measurement data is input into the network, passing sequentially through an encoder, skip connections, and a decoder to obtain the intelligent segmentation result of the measurement plane. In the encoder stage, the accumulated echo data is first downsampled step-by-step using three layers of convolution and pooling operations. Then, block embedding and positional encoding are performed on the obtained feature maps to extract target features, and pixel-level classification of the target and background is completed based on these features. Further, the embedded one-dimensional vector is input into the Transformer module for hybrid encoding. Finally, through skip connections, the encoded feature map is sent to the corresponding decoder layer, where it is concatenated with the data recovered through deconvolution upsampling to recover the corresponding pixel information in the measurement data plane.
[0115] During the nth layer processing of the encoder, given the input features Its output is The intermediate process value is denoted as The calculation of this layer consists of two core steps, expressed as follows:
[0116]
[0117]
[0118] In this module, MSA (Multi-Head Self-Attention Module) represents the multi-head self-attention module, MLP (Multilayer Perceptron Module) represents the multilayer perceptron module, and LN (Layer Normalization) represents the layer normalization operation. Compared to traditional convolutional neural networks, the MSA module projects input features onto multiple parallel attention heads, enabling the network to simultaneously aggregate rich long-range spatial dependencies from different feature subspaces. This is key to TransUNet's ability to effectively capture global information.
[0119] The MSA module works as follows: Assuming the normalization function is softmax, the output vector at the i-th position in the sequence can be represented as:
[0120]
[0121] The query (Q), key (K), and value (V) matrices are formed by combining the input features X with learnable weight matrices. , , Obtained by linear transformation, i.e. , , . This refers to the dimension of vector K. The formula shows that by calculating the correlation between Q and each K, the weight coefficient corresponding to each V can be obtained. Then, V is weighted and summed to obtain the final attention value. Therefore, the essence of the self-attention mechanism is to perform weighted aggregation of V, while Q and K are used to calculate the weights.
[0122] Multi-head self-attention is an extension of this mechanism. It learns multiple independent Q, K, V matrices in parallel and computes self-attention for each:
[0123]
[0124] Here, `head` is a self-attention head. The outputs of all `head`s are concatenated along the channel dimension and fused using a linear projection matrix `W`.
[0125]
[0126] Here, h represents the number of self-attention heads in the multi-head self-attention mechanism. This design allows the model to enhance its attention to the global motion trajectory of the target in the spatial dimension and suppress its response to noise. Simultaneously, by calculating the weight coefficient matrix and the softmax after the dot product, not only are connections established between similar information, but also between different information, thus gaining a more comprehensive understanding of the target's global contextual information. It is important to note that the self-attention mechanism itself cannot perceive the order or positional information of the sequence; therefore, positional encoding is required before inputting the features of the measurement data plane into the Transformer.
[0127] Through the layer-by-layer processing of the encoder, the network ultimately extracts the motion features of weakly maneuvering targets, achieving pixel-based binary classification of the target and background noise. After binarization, the output is a visualized image that clearly identifies the target's location.
[0128] Step 3: Fit the potential trajectory region based on Huber loss M-estimation and extract the starting point, direction of motion, and velocity of the potential trajectory region.
[0129] Suppose that in the binary mask output by the above steps, the k-th candidate connected region contains N pixels. Let this set of pixels be represented as... Each pixel This represents its position in the cumulative image coordinate system.
[0130] Because deep learning segmentation results are irregular at low signal-to-noise ratios, traditional least squares methods, which use squared error loss, are highly susceptible to outliers, causing the fitted line to deviate from the true trajectory. Therefore, this invention introduces M-estimation theory and uses the Huber loss function, which is insensitive to outliers, to construct a regression model.
[0131] Construct a fitted line L, which is formed by a point on the line. and unit direction vector The goal of optimization is to find the optimal parameters. This minimizes the sum of Huber losses of the residuals from all pixels to the line.
[0132] objective function The definition is as follows:
[0133]
[0134] in, For the i-th pixel The fitting residual to line L:
[0135]
[0136] The Huber loss function is defined as a piecewise function:
[0137]
[0138] When the residual When, the loss function is expressed as The norm ensures the fitting accuracy for interior points, when the residuals At that time, the loss function is expressed as follows: Norms limit the influence of edge outlier noise, where This is a preset threshold parameter.
[0139] By solving the above objective function using the iterative reweighted least squares method, the motion direction with optimal noise resistance can be obtained. And the center point C.
[0140] Since the fitted straight line L extends infinitely, the physical range of the trajectory needs to be determined based on the pixel distribution. Subsequently, the projection method is used to determine the geometric endpoints of the trajectory.
[0141] Calculate each pixel in the point set P In the direction of the main axis of motion scalar projection value on :
[0142]
[0143] Iterate through all projected values and extract the minimum value. With the maximum value And based on this, reconstruct the two geometric endpoints of the trajectory. and :
[0144] .
[0145] Step 4: Perform local DP-TBD on the prior information obtained in Step 3.
[0146] Local search space establishment: Let the original measurement data be... For the k-th frame Predict the center position of candidate targets based on a uniform motion model. :
[0147]
[0148] Define the local search region for the k-th frame. For A rectangular area centered at R with a search radius of R:
[0149]
[0150] Where R is the preset search radius, used to address potential errors in parameter extraction during the feature extraction step.
[0151] Local DP-TBD with smooth constraints:
[0152] 1) Initialization: When At that time, for all discrete states, we have:
[0153]
[0154]
[0155] 2) Iterative accumulation: when Then, the following calculations are performed on all discrete states:
[0156]
[0157]
[0158] in, Represents a value function. This represents the detection statistics for a single frame. Indicates the transition relationship between states. express The set of state transitions at time t, To smooth out the penalty coefficient, This represents the Euclidean distance between the current state and the state of the previous frame.
[0159] 3) Threshold decision: when At that time, the following calculations are performed:
[0160]
[0161] The detection threshold is calculated from the false alarm rate and signal-to-noise ratio. If the conditions of the above formula are met, the trajectory is backtracked; otherwise, the trajectory restoration is declared to have failed, and the process ends here.
[0162] 4) Track retracing:
[0163] make Perform the following operations:
[0164]
[0165] Finally, the discrete trajectory of the target is obtained. .
[0166] The invention is verified through simulation experiments and measured data experiments. During the experiments, the measurement plane size was 256×256. There were 3 targets, 25 detection frames, and the motion speed and direction were randomly generated in the scene. Monte Carlo simulation was used to verify the performance of the proposed detector, with a false alarm rate set at [value missing]. Monte Carlo times First, the TransUNet network is used to segment the image, and features such as the starting position and direction of the candidate trajectory are extracted. Then, the trajectory is reconstructed using local DP-TBD search. Finally, the detection performance of this method is evaluated by comparing the trajectory detection probability and computational complexity with other methods.
[0167] Figure 2 The results of segmentation using the TransUNet network are shown under signal-to-noise ratio conditions of 3dB, 6dB, and 10dB. Figure 2In the figure, (a) is the measurement space plane, (b) is the TransUNet network segmentation result, and (c) is the mask. It can be seen from the figure that in the low signal-to-noise ratio range, the network segmentation has false positives and false negatives.
[0168] Figure 3 To demonstrate the performance advantages of the proposed method compared to HF-TBD under different signal-to-noise ratio conditions, the effective detection probability is set as follows: This is represented by the ratio of the number of successfully reconstructed points in the trajectory to the total number of trajectory points. The condition for successful trajectory reconstruction is that the distance between the detected waypoint and the actual waypoint is less than 2 units, and... .
[0169] Figure 4 To compare the complexity of this invention with that of the traditional DP-TBD algorithm, under the same parameter conditions and operating environment, after using the local feature extraction of the TransUNet network, the floating-point running time has a considerable speedup compared to the traditional DP-TBD running time, and the speedup increases with the number of running frames.
[0170] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A multi-target detection pre-tracking method based on TransUnet segmentation guidance, characterized in that, Includes the following steps: (1) Construct a motion model and use maximum value accumulation to perform non-coherent superposition of multi-frame measurement data; (2) The TransUnet network is used to segment the measurement data plane, extract global context features, and generate candidate regions for potential target trajectories; (3) Based on Huber loss, M-estimation fits the potential target trajectory and extracts the starting point, direction of motion, and velocity of the potential target trajectory in parallel; (4) Using the extracted position, velocity and direction parameters as prior guidance, a local state space under kinematic constraints is constructed based on multi-frame measurement data; the trajectories of all candidate regions are estimated in detail using dynamic programming before detection tracking algorithm within the local state space. (5) Perform trajectory backtracking and threshold detection on the estimation results of step (4).
2. The multi-target detection pre-tracking method based on TransUnet segmentation guidance according to claim 1, characterized in that, The process of constructing the motion model in step (1) is as follows: The target moving in space The motion state at any given time consists of motion parameters including the target's position and velocity, where K is the total number of frames processed in a batch by the tracking algorithm. ; Its state variables are: ； in, , The position of the target point in the X and Y dimensions. , These are the velocities of the target point in the X and Y directions, respectively. Consider a moving target in the XY plane. Its motion state evolution follows a one-dimensional Markov process, and its motion model is expressed as: ； Where Q represents process noise. Representing the inter-frame interval, when the target is moving at a constant velocity in a straight line in the XY plane, the above formula is: ； Where F is the state transition matrix: ； in, It is a second-order identity matrix. Let the Kronecker product be represented; then the track of a target in a batch of K frames is represented as: ； For the motion model, each measurement unit is a two-dimensional pixel plane, assuming the size of this pixel plane is... Then the measurement plane at time k Represented as a The matrix: ； Use respectively and Let represent the scenarios where the target exists and the scenario where it does not exist, respectively. The statistical characteristics of each resolution unit are given by the following formula: ； in, It is independent and identically distributed zero-mean complex Gaussian white noise. For constant amplitude and phase in A complex random variable whose internal distribution follows a uniform distribution.
3. The multi-target detection pre-tracking method based on TransUnet segmentation guidance according to claim 1, characterized in that, The process of non-coherently superimposing multi-frame measurement data using maximum value accumulation in step (1) is as follows: The measurement data plane is represented as follows: ； in pixel unit Represented as: 。 4. The multi-target detection pre-tracking method based on TransUnet segmentation guidance according to claim 1, characterized in that, The implementation process of step (2) is as follows: The preprocessed measurement data is input into the TransUnet network, passing through an encoder, skip connections, and a decoder in sequence to obtain the intelligent segmentation result of the measurement plane. In the encoder stage, the accumulated echo data is first downsampled stepwise through three layers of convolution and pooling operations. Then, the obtained feature map is subjected to block embedding and position encoding to extract target features and perform pixel-level classification of the target and background based on the features. Further, the embedded one-dimensional vector is input into the Transformer module for hybrid encoding. Finally, through skip connections, the encoded feature map is sent to the corresponding decoder layer and concatenated with the data recovered by deconvolution upsampling to recover the corresponding pixel information in the measurement data plane. During the nth layer processing of the encoder, given the input features Its output is The intermediate process value is denoted as The calculation of this layer consists of two core steps, expressed as follows: ； ； Among them, MSA stands for Multi-head Self-Attention Module, MLP stands for Multilayer Perceptron Module, and LN represents the normalization operation of the layer.
5. The multi-target detection pre-tracking method based on TransUnet segmentation guidance according to claim 4, characterized in that, The implementation process of the multi-head self-attention module (MSA) is as follows: If the normalization function is softmax, then the output vector at the i-th position in the sequence is represented as: ； The query Q, key K, and value V matrices are formed by combining the input features X with learnable weight matrices. , , Obtained by linear transformation, i.e. , , ; It is the dimension of vector K; by calculating the correlation between Q and each K, the weight coefficient corresponding to each V is obtained, and then the V are weighted and summed to obtain the final attention value; Multi-head self-attention learns multiple independent Q, K, V matrices in parallel, and calculates self-attention for each matrix separately: ； Here, `head` is a self-attention head. The outputs of all `head`s are concatenated along the channel dimension and fused using a linear projection matrix `W`. ； Where h is the number of self-attention heads in multi-head self-attention.
6. The multi-target detection pre-tracking method based on TransUnet segmentation guidance according to claim 1, characterized in that, The implementation process of step (3) is as follows: Suppose that in the output binary mask, the k-th candidate connected region contains N pixels, and this set of pixels is represented as... Each pixel This represents its position in the cumulative image coordinate system; Introducing the M-estimation theory, a regression model is constructed using the Huber loss function, which is insensitive to outliers: a fitted straight line L is constructed, which is formed by a point on the line... and unit direction vector The goal of optimization is to find the optimal parameters. This minimizes the sum of Huber losses of the residuals of all pixels to the line; objective function The definition is as follows: ； in, For the i-th pixel The fitting residual to line L: ； The Huber loss function is defined as a piecewise function: ； When the residual When, the loss function is expressed as The norm ensures the fitting accuracy for interior points, when the residuals At that time, the loss function is expressed as Norms limit the impact of edge outlier noise, where This is a preset threshold parameter; The objective function is solved by iterative reweighted least squares method, thus obtaining the direction of motion. and center point C; Calculate each pixel in the point set P In the direction of the main axis of motion scalar projection value on : ； Iterate through all projected values and extract the minimum value. With the maximum value And based on this, reconstruct the two geometric endpoints of the trajectory. and : 。 7. The multi-target detection pre-tracking method based on TransUnet segmentation guidance according to claim 1, characterized in that, The implementation process of step (4) is as follows: Establish a local search space: Let the original measurement data be... For the k-th frame, the center position of the candidate target is predicted based on the uniform motion model. : ； Define the local search region for the k-th frame. For A rectangular area centered at R with a search radius of R: ； Where R is the preset search radius, used to deal with possible errors in parameter extraction during the feature extraction step; when At that time, for all discrete states, we have: ； ； when Then, the following calculations are performed for all discrete states: ； ； in, Represents a value function. This represents the detection statistics for a single frame. Indicates the transition relationship between states. express The set of state transitions at time t, To smooth out the penalty coefficient, This represents the Euclidean distance between the current state and the state of the previous frame.
8. The multi-target detection pre-tracking method based on TransUnet segmentation guidance according to claim 1, characterized in that, The implementation process of step (5) is as follows: Threshold decision: when At that time, the following calculations are performed: ； in, The detection threshold is calculated from the false alarm rate and signal-to-noise ratio. If the conditions of the above formula are met, the trajectory is backtracked; if not, the trajectory restoration is declared to have failed. Flight path retracing: Order Perform the following operations: ； Obtain the discrete trajectory of the target .