A non-cooperative bistatic radar detection forward tracking algorithm
By applying nuclear correlation filtering and deep learning techniques to a non-cooperative bistatic radar, the drift and loss problems in weak target tracking were solved, and a stable pre-detection tracking effect was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NAVAL AVIATION UNIV
- Filing Date
- 2023-06-21
- Publication Date
- 2026-06-02
AI Technical Summary
Non-cooperative bistatic radars face challenges such as weak echo signals and unstable target detection under low signal-to-noise ratio conditions. Traditional pre-detection tracking methods involve large amounts of data processing and rely on motion models. KCF tracking frames are susceptible to clutter interference, leading to drift and loss.
Kernel correlation filtering is used for pre-detection tracking, combined with deep learning for trajectory prediction, using peak sidelobe ratio as the evaluation index for tracking loss, and neural network is used to correct the tracking box position, reducing data processing pressure and avoiding dependence on motion models.
It improves the stability of non-cooperative bistatic radar in tracking weak targets, reduces tracking frame drift and loss, and achieves efficient target detection and tracking.
Smart Images
Figure CN116736291B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a KCF-TBD method based on non-cooperative bistatic radar, belonging to the field of weak target tracking technology of non-cooperative bistatic radar. Background Technology
[0002] Non-cooperative bistatic radar is a passive target detection system that uses a non-cooperative radar as the radiation source and operates based on the bistatic radar principle. Compared to traditional active monostatic radar, non-cooperative bistatic radar has advantages such as better concealment, lower cost, and stronger "four-proof" capabilities (anti-theft, anti-reflective, anti-interference, and anti-reflective properties). Its receiving system works by using the direct wave signal emitted by the non-cooperative radar radiation source as a reference to detect and analyze the target's scattered echo signal, thereby achieving target location and tracking. However, due to the uncertainty of the radiation source and the limitations of the bistatic geometric model, non-cooperative bistatic radar faces problems such as weak echo signals, low target signal-to-noise ratio, and unstable target detection between radar scanning frames during detection.
[0003] To improve the detection performance of a system under low signal-to-noise ratio conditions, the simplest approach is to lower the detection threshold. However, excessively lowering the threshold can lead to clutter and generate numerous false alarms. In this case, the traditional Detect-Before-Track (DBT) method struggles to achieve reliable target detection and stable tracking. Therefore, the Track-Before-Detect (TBD) method was proposed. Unlike DBT, TBD does not perform threshold detection processing within a single frame or sets a small threshold. Instead, it digitizes and stores the radar echo data, and after joint processing of multiple frames, announces the detection result and simultaneously estimates the target trajectory. However, traditional TBD methods suffer from large data processing volumes and reliance on motion prior models. Therefore, reducing the computational load of data processing and making the tracking process independent of motion models is crucial. With the rapid development of computer vision technology, kernelized correlation filters (KCF) have been widely applied in target tracking due to their high efficiency, speed, and strong target recognition capabilities. KCF treats target tracking as a similarity matching problem based on target templates, essentially equivalent to the value function accumulation idea in traditional pre-detection tracking. Its main matching method uses Fourier transform to calculate the convolution between the target template and the current image, thus obtaining the target's position in the current image. This method does not require a prior model of motion and can improve computational efficiency by utilizing the diagonalization property of circumcircle matrices. However, typical KCF tracking often suffers from clutter interference, leading to bounding box drift and loss.
[0004] Therefore, a pre-detection tracking method that can be used in non-cooperative bistatic radar scenarios is needed. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a non-cooperative bistatic radar pre-detection tracking method. This method enables the detection and tracking of multiple weak targets without relying on a motion model. Furthermore, it addresses the issues of unstable target acquisition between frames in passive bistatic radar during actual scanning, where the KCF tracking frame is susceptible to strong clutter and system errors, leading to tracking frame drift and loss. This invention utilizes deep learning for trajectory prediction, effectively correcting the KCF tracking frame position and improving the stability of the tracking process. The invention also provides a non-cooperative bistatic radar detection system, primarily composed of a broadband electromagnetic signal real-time measurement and analysis instrument, a directional grid antenna, a rod-shaped omnidirectional antenna, and a data processing and display PC.
[0006] The present invention provides a non-cooperative bistatic radar detection-pre-tracking algorithm, characterized by the following steps:
[0007] Step 1: Receive the direct wave reference signal from the non-cooperative radar radiation source and the target echo radio frequency analog signal through a directional grid antenna and a rod-shaped omnidirectional antenna. After receiving and processing by the broadband electromagnetic signal real-time measurement and analysis instrument, the reference signal and target echo IQ baseband data are output. Finally, the data is sent to the data processing and display PC for pre-detection tracking algorithm processing.
[0008] Step 2: Cross-correlate each received frame of non-cooperative bistatic radar signal with the reference signal to obtain the corresponding range-azimuth matrix;
[0009] Step 3: Perform low-threshold CFAR detection on the range-azimuth matrix of the first and second frames. The purpose is to detect moving and stationary targets with weak echo signals as much as possible. During detection, points that pass the threshold are recorded as 1 and points that do not pass the threshold are recorded as 0, so as to binarize the range-azimuth matrix.
[0010] Step 4: Use the OpenCV image contour extraction function to select the irregular target with a rectangular box, and obtain the first tracking box set and the second tracking box set respectively;
[0011] Step 5: Perform an IOU (Intersection over Union) decision on the first and second bounding box sets to filter out moving targets. The IOU calculation formula is as follows: Where IntersectionArea is the area of the intersection of the two regions, Area1 and Area2 represent the areas of the two regions to be compared, and IOU T It is the threshold for the decision, with an overlap greater than 0 and less than IOU.T Those that are moving targets are identified as moving targets, while those that are stationary targets are identified as stationary targets.
[0012] Step 6: After the IOU decision in the first two frames, a new set of tracking boxes is obtained, which represent the moving targets within them;
[0013] Step 7: Use the obtained new tracking box set to train the filter, assuming the objective function is f(x) i )=ω T x i ω is the weight coefficient in the mapping relationship. A regularization term ν is added to prevent overfitting. The learning objective function is...
[0014] Step 8: To further improve learning performance, x i and y i The relationship between them Mapping to a higher dimension At this point, the variable being optimized changes from α to ω. Therefore, after differentiating with respect to α, the solution to the ridge regression problem is α = (K + νI). -1 y, where K is the kernel matrix in the kernel space, I is the n*n identity matrix, and α = [α1, α2, ..., α n ] T y = [y1, y2, ..., y n ] T ;
[0015] Step 9: Since K is a cyclic matrix, utilize the property of diagonalization of cyclic matrices. achievable Simultaneously, using the weight α learned in step 7 for the first frame, the corresponding matrix for the k-th frame is... in and Representing the Fourier transform and its inverse transform respectively, Substituting, we can obtain By repeating steps 6-8, the target can be tracked.
[0016] Step 10: The response matrix obtained in Step 9 may experience drift and tracking loss due to the inherent properties of non-cooperative bistatic radar. Therefore, the peak signal-to-noise ratio (PSR) is selected as the evaluation metric for tracking loss, defined as follows: Where g max It is the response matrix y K The maximum value is used to select a rectangular area around the maximum value as the target's peak window. The area inside the tracking box and outside the peak window is called the sidelobe window. s and σ sThese are the average and standard deviation of the sidelobe windows. The average PSR of three consecutive frames is calculated as a threshold. If the PSR calculated in subsequent frames is less than this threshold by a certain range, the tracking box is considered lost.
[0017] The PSR in step 10 can be used as a value function of the pre-detection tracking technology of this patent, as explained below:
[0018] 1) The core idea of traditional dynamic programming for pre-detection tracking is to search all possible paths in discrete space according to the prior model of the target motion and estimate the optimal target path. Here, λ is introduced as a value function, with the aim of finding the target path sequence X that maximizes the value function λ among all possible combinations of target paths. 1:K =(x1,…,x K Therefore, the target track sequence can be represented as...
[0019] 2) Set a threshold γ, when a certain state sequence When the accumulated value function exceeds a threshold, a track is detected, and track backtracking is initiated.
[0020] 3) Therefore, for KCF-TBD, the proposition of the cumulative maximum trajectory generated by the value function of the KCF tracker can be similarly put forward. Assume... If the KCF tracker's state sequence is within K frames, then the cumulative sum of this state sequence is the maximum cumulative sum across all physically feasible paths.
[0021] 4) For each KCF tracker, it searches through all possible expansion cells of the tracking frame to obtain the response matrix y. K The tracker ultimately selects the matrix with the largest response for training and re-detection in the next frame. During this process, the tracker uses the maximum response as the result of target tracking at each step. Traditional dynamic programming pre-detection tracking uses the maximum accumulated value of the value function along the entire path to represent the track, while the pre-detection tracking method proposed in this patent uses the maximum response (value function) between frames instead of the original accumulation of the value function. Considering this from the perspective of inequalities, we can obtain...
[0022] 5) Regarding (10) And g max It is the response matrix y K The maximum value, which means that the PSR when the KCF tracker detects a target can be used as a value function. The equation in equation 4 can be rewritten as follows:
[0023] Step 11: When the PSR decision module determines that the tracking box is lost, the prediction correction module is activated to adjust the tracking box;
[0024] Furthermore, the prediction correction module in step 11 is obtained by training a neural network; the specific network structure is shown in the appendix. Figure 2 The network mainly consists of three parts: a fully connected input network, an LSTM network, and a fully connected output network. The input features are the positions of the tracking boxes in the range-azimuth plane for s consecutive frames, and the output nodes are the positions of the tracking boxes predicted by the network model in the (s+1)th frame. The training data is trajectory data generated from range-azimuth simulation, and the models used are of two types: uniform velocity and uniform acceleration. The actual position of the KCF tracking box during the tracking process is represented by the position of the pixel, so the simulated trajectory needs to be converted from latitude and longitude to a Cartesian coordinate system, and then converted to the coordinates of the pixels in the range-azimuth plane based on trigonometric relationships.
[0025] Step 12: After the prediction correction module, the corrected tracking box is obtained, and a new tracker template is trained using this as a template to re-track;
[0026] Step 13: Repeat steps 10-12 until all of them form a continuous track.
[0027] The non-cooperative bistatic radar pre-detection tracking method of the present invention has the following beneficial effects:
[0028] This invention proposes a pre-detection tracking method based on non-cooperative bistatic radar, utilizing kernel correlation filter technology from computer vision for tracking. To address issues such as tracking box drift and loss, this method employs peak-to-side-lobe ratio as an evaluation metric for tracking loss and applies deep learning methods to adjust the tracking box. Compared to traditional pre-detection tracking methods, this method has lower data processing requirements and does not require a prior model of motion, thus achieving higher stability in weak target tracking under non-cooperative bistatic radar scenarios. Attached Figure Description
[0029] Figure 1 Flowchart of a non-cooperative bistatic radar pre-detection tracking method;
[0030] Figure 2 This is a diagram of the model network structure. Detailed Implementation
[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0032] Example 1
[0033] This embodiment of a non-cooperative bistatic radar pre-detection tracking method specifically includes the following steps:
[0034] Step 1: Receive the direct wave reference signal from the non-cooperative radar radiation source and the target echo radio frequency analog signal through a directional grid antenna and a rod-shaped omnidirectional antenna. After receiving and processing by the broadband electromagnetic signal real-time measurement and analysis instrument, the reference signal and target echo IQ baseband data are output. Finally, the data is sent to the data processing and display PC for pre-detection tracking algorithm processing.
[0035] Step 2: Cross-correlate each received frame of non-cooperative bistatic radar signal with the reference signal to obtain the corresponding range-azimuth matrix;
[0036] Step 3: Perform low-threshold CFAR detection on the range-azimuth matrix of the first and second frames. The purpose is to detect moving and stationary targets with weak echo signals as much as possible. During detection, points that exceed the threshold are marked as 1, and points that do not exceed the threshold are marked as 0, thus binarizing the range-azimuth matrix. The specific parameters of the low-threshold CFAR detector are set as follows: number of reference cells 32; number of guard cells 16; scaling factor 1.3.
[0037] Step 4: Using the OpenCV image contour extraction function, irregular targets can be selected with rectangular boxes to obtain the first tracking box set and the second tracking box set respectively;
[0038] Step 5: Perform an IOU (Intersection over Union) decision on the first and second bounding box sets to filter out moving targets. The formula for calculating IOU is... Where Intersection Area is the area of the intersection of two regions, and Area1 and Area2 represent the areas of the two regions to be compared. T It is the threshold for the decision, with an overlap greater than 0 and less than IOU. T Those that are moving targets are identified as moving targets, while those that are stationary targets are identified as stationary targets.
[0039] Step 6: After the IOU decision in the first two frames, a new set of tracking boxes is obtained, which represent the moving targets within them;
[0040] Step 7: Utilize the new set of tracking boxes and extract the image features within these tracking boxes, convert them to the frequency domain to obtain the target template, and simultaneously learn the initial filter coefficients.
[0041] Step 8: At each time point, select a search window, extract the image features within the window, transform the features, and then use Fourier transform to map the features within the search window to the frequency domain to obtain the response map of the search window.
[0042] Step 9: In the frequency domain, convolve the target region template and the response map of the search window to obtain the similarity distribution map;
[0043] Step 10: Find the location of the maximum response value in the similarity distribution map, which is the new location of the target. Simultaneously, use this new location as the center of the target region to extract a new target area, and extract the image features of this region again. Finally, update the filter coefficients by performing linear regression on the newly extracted image features to make the filter more accurately adapt to changes in the target's appearance.
[0044] Step 11: Calculate the peak sidelobe ratio (PSR) each time a new target location is obtained. Where g max It is the response matrix y K The maximum value is used to select a rectangular area around the maximum value as the target's peak window. The area inside the tracking box and outside the peak window is called the sidelobe window. s and σ s These are the average and standard deviation of the sidelobe windows. The average PSR of three consecutive frames is calculated as a threshold. If the PSR calculated in subsequent frames is less than this threshold by a certain range, the tracking box is considered lost.
[0045] Step 12: When the PSR decision module determines that the tracking box is lost, the prediction correction module is activated to adjust the tracking box;
[0046] The prediction and correction module in step 12 is obtained by training a neural network. See the appendix for the specific network structure. Figure 2 The network mainly consists of three parts: a fully connected input network, an LSTM network, and a fully connected output network. The input features are the positions of the tracking boxes in the range-azimuth plane for s consecutive frames, and the output nodes are the positions of the tracking boxes predicted by the network model in the (s+1)th frame. The training data is trajectory data generated from range-azimuth simulation, and the models used are of two types: uniform velocity and uniform acceleration. The actual position of the KCF tracking box during the tracking process is represented by the position of the pixel, so the simulated trajectory needs to be converted from latitude and longitude to a Cartesian coordinate system, and then converted to the coordinates of the pixels in the range-azimuth plane based on trigonometric relationships.
[0047] Specifically, step 12 can be further subdivided into the following steps:
[0048] Step 12.1: Treat each track in the training data as a sample. Assume the length of the sequence is sq_len. Each point in the sequence can be represented as p = {x, y}, where x and y are the target positions obtained after latitude and longitude coordinate conversion. This position is the result of conversion with the non-cooperative bistatic radar receiver as the center.
[0049] Step 12.2: The entire sequence can be represented as Where i represents the dimension of the input network Batch_size; This represents the input sequence information, with a length of sq_len; This represents the track position at time sq_len+1, which is equivalent to the true value that needs to be predicted;
[0050] Step 12.3: Normalize the data of size Batch_size;
[0051] Step 12.4: Feed the normalized data into the network for training. The dimension of each batch of data is Batch_size×sq_len×input_data, where the dimension of input_data is the dimension of the track points.
[0052] Step 12.5: Calculate the loss function for each batch of data, perform backpropagation, and update the network weights using the SGD optimizer;
[0053] Step 12.6: After each batch of training data is completed, test it and calculate its accuracy;
[0054] Step 12.7: Save the weights of the model with the highest accuracy for use in correcting the tracking boxes;
[0055] Step 13: After the prediction correction module, the corrected tracking box is obtained, and a new tracker template is trained using this as a template to re-track;
[0056] Step 14: Repeat steps 10-11 until all of them form a continuous track.
[0057] This invention proposes a pre-detection tracking method based on non-cooperative bistatic radar, utilizing kernel correlation filter technology from computer vision for tracking. To address issues such as tracking box drift and loss, this method employs peak-to-side-lobe ratio as an evaluation metric for tracking loss and applies deep learning methods to adjust the tracking box. Compared to traditional pre-detection tracking methods, this method has lower data processing requirements and does not require a prior model of motion, thus achieving higher stability in weak target tracking under non-cooperative bistatic radar scenarios.
Claims
1. A non-cooperative bistatic radar pre-detection tracking algorithm, characterized in that... Includes the following steps: Step 1: Receive the direct wave reference signal from the non-cooperative radar radiation source and the target echo RF analog signal through a directional grid antenna and a rod-shaped omnidirectional antenna. After receiving and processing by a broadband electromagnetic signal real-time measurement and analysis instrument, the reference signal and target echo IQ baseband data are output. Finally, the data is sent to a data processing and display PC for pre-detection tracking algorithm processing. Step 2: Cross-correlate each received frame of non-cooperative bistatic radar signal with the reference signal to obtain the corresponding range-azimuth matrix; Step 3: Perform low-threshold CFAR detection on the range-azimuth matrix of the first and second frames. During detection, all points that pass the threshold are recorded as 1, and all points that do not pass the threshold are recorded as 0, so as to binarize the range-azimuth matrix. Step 4: Use OpenCV's image contour extraction function to select irregular targets with rectangular boxes, and obtain the first tracking box set and the second tracking box set respectively; Step 5: Perform an IOU (Intersection over Union) calculation on the first and second bounding box sets to filter out moving targets. The formula for calculating IOU is shown in the formula below. ； Intersection Area is the area of the intersection of two regions. Area1 and Area2 represent the areas of the two regions to be compared. It is the threshold for judgment, with an overlap greater than 0 and less than 1. Those that are moving targets are identified as moving targets, while those that are stationary targets are identified as stationary targets. Step 6: After the IOU decision in the first two frames, a new set of tracking boxes is obtained, which represent the moving targets within them; Step 7: Use the obtained new tracking box set to train the filter, assuming the objective function is... ω is the weight coefficient in the mapping relationship, and a regularization term is added. To prevent overfitting, the objective function for learning is... ; Step 8: To further improve learning performance, and The relationship between them Mapping to a higher dimension Now, the variable to be optimized changes from α to ω. Therefore, after differentiating with respect to α, the solution to the ridge regression problem is... Where K is the kernel matrix in the kernel space. It is an n*n identity matrix. , ; Step 9: Since K is a cyclic matrix, utilize the property of diagonalization of cyclic matrices. achievable Simultaneously, using the weight α learned in step 7 for the first frame, the corresponding matrix for the k-th frame is... ,in and Representing the Fourier transform and its inverse transform respectively, Substituting, we can obtain By repeating steps 7-9, target tracking can be achieved. Step 10: The response matrix obtained in Step 9 may experience drift and tracking loss due to the inherent properties of non-cooperative bistatic radar. Peak sidelobe ratio (PSR) is selected as the evaluation metric for tracking loss, defined as follows: ,in It is the response matrix The maximum value is used to select a rectangular area around the maximum value as the target's peak window. The area inside the tracking box and outside the peak window is called the side lobe window. and It is the average value and standard deviation of the sidelobe window. The average value of the PSR of three consecutive frames is used as the threshold. When the PSR calculated in subsequent frames is less than this threshold by a certain range, the tracking box is determined to be lost. Step 11: When the PSR decision module determines that the tracking box is lost, the prediction correction module is activated to adjust the tracking box; Step 12: After the prediction correction module, the corrected tracking box is obtained, and a new tracker template is trained using this as a template to re-track; Step 13: Repeat steps 10-12 until all of them form a continuous track.
2. The non-cooperative bistatic radar pre-detection tracking algorithm according to claim 1, characterized in that... In step 10, PSR serves as the value function of the pre-detection tracking technique. The specific explanation is as follows: 1) The core idea of traditional dynamic programming pre-detection tracking is to search for all possible trajectories in discrete space according to the prior model of the target motion, and estimate the optimal target trajectory. As a value function, the goal is to find the value function among all possible combinations of target trajectories. Target track sequence that reaches the maximum value Therefore, the target track sequence is represented as ; 2) Set a threshold When a certain state sequence When the accumulated value function exceeds a threshold, a track is detected, and track backtracking is initiated. ; 3) Therefore, for KCF-TBD, similarly, we propose the proposition that the KCF tracker generates the cumulative maximum trajectory of the value function, assuming... If the KCF tracker's state sequence is within K frames, then the cumulative sum of this state sequence is the maximum cumulative sum across all physically feasible paths. ; 4) For each KCF tracker, it searches through all possible expansion cells of the tracking frame to obtain the response matrix. The tracker ultimately selects the matrix with the largest response for training and re-detection in the next frame. During this process, the tracker uses the maximum response as the result of target tracking each time. Traditional dynamic programming pre-detection tracking uses the maximum accumulated value of the value function along the entire path to represent the track, while the proposed pre-detection tracking method uses the maximum response between frames instead of the original accumulated value function. From the perspective of inequalities, we can obtain... ; 5) Regarding step 10 , It is the response matrix The maximum value of the PSR when the KCF tracker detects the target is used as a value function. The expression in 4) can be rewritten as .
3. The non-cooperative bistatic radar pre-detection tracking algorithm according to claim 1, characterized in that... The prediction correction module in step 11 is trained by a neural network. The network mainly consists of three parts: an input fully connected network, an LSTM network, and an output fully connected network. The input features of the network are the positions of the distance-azimuth plane tracking boxes in s consecutive frames, and the output nodes are the predictions made by the network model. The position of the frame tracking box is trained using trajectory data generated from range-azimuth simulation. The models used are of two types: uniform velocity and uniform acceleration. The actual position of the KCF tracking box during the tracking process is represented by the position of the pixel. Therefore, the simulated trajectory needs to be converted from latitude and longitude to a Cartesian coordinate system, and then converted into the coordinates of the pixel points in the range-azimuth plane based on trigonometric relationships.