Through-the-wall radar target tracking method based on scale self-adaption-rotation kernel correlation filtering
By adopting the scale adaptive-rotating core-related filtering method in the target tracking of the wall-through radar, the tracking difficulties caused by the change in the target image scale and angle are solved, and high-precision target tracking is achieved.
Patent Information
- Application Number
- CN202510360333.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-25
- Publication Date
- 2025-06-27
AI Technical Summary
The existing wall-through radar target tracking method is difficult to effectively track the position changes of the target image when the scale and angle change, resulting in difficulty in tracking.
The wall-through radar target tracking method based on scale adaptive-rotating core correlation filter is adopted. By extracting the scale and angle information of the target area on the radar image domain, performing scale and rotation transformation, and updating the correlation filter, accurately estimating and tracking the target position is achieved.
This method can realize automatic tracking of the target area without target measurement values, with better tracking accuracy, and is especially suitable for target images with changes in scale and rotation angle.
Smart Images
Figure CN120214784A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of through-wall radar tracking, and particularly relates to a through-wall radar target tracking method based on scale adaptive-rotational kernel correlation filtering. Background Art
[0002] With the development of urbanization, the urban environment is becoming increasingly complex. The occlusion between buildings causes some dangerous personnel to be in a non-line-of-sight scenario, posing a safety hazard. Through-the-Wall Radar (TWR) detects targets by transmitting electromagnetic waves through buildings, and has the advantages of good mobility and strong adaptability to complex environments, and has been widely used in many fields. The moving target tracking technology inside buildings provides key technical support for urban warfare, anti-terrorism rescue, etc.
[0003] Classical filtering algorithms have been widely applied in the field of radar target tracking due to their simple principles. By modeling the motion of the target, the Kalman filtering algorithm can track simple linear motion targets. The interacting multiple model algorithm can track maneuverable human targets with complex and random motions. Particle filtering is used in nonlinear systems to estimate the target state by integrating particle samples. The measurement value is the core for the above algorithms to achieve target tracking, but it is usually difficult to obtain in actual tracking scenarios.
[0004] The mean shift algorithm can automatically search for the target area matching the target template by extracting the target image features to establish the target template. The kernel correlation filtering algorithm uses the target area to train the correlation filter, which can effectively separate the target from the background. We proposed a rotational kernel correlation filtering algorithm in previous research, which improved the adaptability of the tracking algorithm to rotated images by considering the angle change of the image when extracting the candidate area and training samples. Then, the position and attitude changes of the target during the movement will not only cause the angle change of its radar image, but also cause the scale change of the target image, resulting in tracking difficulties. Summary of the Invention
[0005] To solve the above problems, the present invention proposes a through-wall radar target tracking algorithm based on scale adaptive-rotational kernel correlation filtering, which can robustly track the target image with simultaneous scale and angle changes, and specifically includes the following operation steps:
[0006] Step 1: Extract the initial target area on the initial frame radar image and calculate its scale and angle information, and then set the ideal response map G;
[0007] Step 2: Extract the candidate target area according to the geometric information of the target in the previous frame, and perform scale and angle transformation to obtain the candidate target area Z;
[0008] Step 3: Perform a correlation operation on the candidate target region Z using the filter W to obtain the response value f(Z), where the position corresponding to the maximum response value is the target position z in the current frame;
[0009] Step 4: Taking the target position z as the center, calculate the scale and angle information of the target region, extract the target region on the radar image, and perform scale and rotation transformation to obtain the training sample M;
[0010] Step 5: Update the filter W according to the new training sample.
[0011] Further, in Step 1, the calculation method for the size and angle information of the initial target region is as follows:
[0012] Select the maximum pixel value within the initial target region as the initial target position z0, and then, with z0 as the center, extract an image block P containing the target region from the radar image. Use the image moment method to calculate the size and angle information of the initial target region on P:
[0013]
[0014]
[0015] where P bina (·) represents a binary image, θ > 0 indicates that the target rotates clockwise, and vice versa indicates counterclockwise rotation.
[0016] Further, in Step 1, according to the width and height of the initial target region, set the ideal response map G as:
[0017]
[0018] where (x0, y0) is the center position of the response map, and σ is the corresponding variance on the x-axis and y-axis,
[0019] Further, in Step 2, the method for obtaining the candidate target region Z is as follows:
[0020] 1. In the k-th frame, with the target position z k-1 of the previous frame as the center, extract an image block on the radar image of the current frame. Then, according to the rotation angle θ k-1 of the target image in the previous frame, perform rotation:
[0021]
[0022] 2. According to the width (N x ) k-1 and height (N y) k-1 , extract the candidate target region centered on z from the rotated image slice k-1
[0023] 3. Perform scale transformation on the candidate target region and re - fix its size to the initial target region size:
[0024]
[0025] Furthermore, in step 3, the calculation method of the response value f(Z) is as follows:
[0026] According to the candidate target region Z after scale and rotation transformation k and the filter W trained in the previous frame k-1 , calculate the response value:
[0027]
[0028] where F(·) represents the Fourier transform, ⊙ represents the Hadamard product, and α k-1 is the filter parameter of the previous frame, and its calculation formula is:
[0029]
[0030] Furthermore, in step 3, the position with the maximum response value is the target position z at the current moment k .
[0031] Furthermore, in step 4, the acquisition method of the training sample M is as follows:
[0032] 1. Centered on the target position z at the current moment k , extract a sub - image containing the target region Then, according to step 1, calculate the length a of the major axis, k the length b of the minor axis k and the rotation angle θ of the elliptical target region; rotate the sub - image :
[0033]
[0034] 2. Centered on z k , with a width of 2a k and a height of 2b k , extract a rectangular target region as the training sample;
[0035] 3. Then, perform scale transformation on the training samples and fix their sizes back to the initial target region sizes:
[0036]
[0037] Furthermore, in step 5, the update method for the filter W is as follows:
[0038] Update the parameter α of the filter W k as follows:
[0039]
[0040] where M k is the new training sample, G is the ideal response map, * represents complex conjugate, and the calculation formula for k is:
[0041]
[0042] Beneficial effects:
[0043] 1. The present invention provides a through-wall radar target tracking method based on scale adaptive-rotation kernel correlation filtering, which is different from traditional target tracking methods. In this method, the target image is directly tracked in the radar image domain without the need for motion modeling. Simulation experiments and actual measurement experiments show that, compared with other methods, the present invention can achieve automatic tracking of the target area without target measurement values, and it is an effective through-wall radar target tracking method.
[0044] 2. The present invention provides a through-wall radar target tracking method based on scale adaptive-rotation kernel correlation filtering, which is different from traditional target tracking methods. In this method, first, use the estimated size and angle information of the target image to perform scale and rotation transformations on the training samples and update the correlation filter. Then, extract the candidate target area according to this information, perform scale and rotation transformations, and perform correlation operations with the filter to achieve target position estimation. Simulation experiments and actual measurement experiments show that, compared with other methods, the present invention has better tracking accuracy for targets with scale and rotation angle changes, and it is an effective through-wall radar target tracking method. Brief Description of the Drawings
[0045] Figure 1 is a simulation scenario diagram of the method of the present invention;
[0046] Figure 2It is the result diagram of processing simulation data by different methods. Among them, (a) is the result diagram of processing simulation data by using the mean shift algorithm, (b) is the result diagram of processing simulation data by using the shape-adaptive mean shift algorithm, (c) is the result diagram of processing simulation data by using the kernel correlation filtering algorithm, (d) is the result diagram of processing simulation data by using the rotational kernel correlation filtering algorithm, and (e) is the result diagram of processing simulation data by using the method of the present invention;
[0047] Figure 3 It is the position error diagram of processing simulation data by different methods;
[0048] Figure 4 It is the experimental scenario diagram of the method of the present invention and the radar array distribution diagram;
[0049] Figure 5 It is the result diagram of processing measured data by different methods. Among them, (a) is the result diagram of processing measured data by using the mean shift algorithm, (b) is the result diagram of processing measured data by using the shape-adaptive mean shift algorithm, (c) is the result of processing measured data by using the kernel correlation filtering algorithm, (d) is the result diagram of processing measured data by using the rotational kernel correlation filtering algorithm, and (e) is the result diagram of processing measured data by using the method of the present invention.
[0050] Specific implementation process
[0051] The object of the present invention is to solve the problem that the existing methods have poor tracking effect on target images with scale and angle changes, and a radar image domain tracking method based on scale-adaptive - rotational kernel correlation filtering is proposed. The present invention will be described in detail below with reference to the accompanying drawings and embodiments, including the following steps:
[0052] Step 1: Extract the initial target area on the initial frame radar image and set the ideal response map G:
[0053] Image formation is performed according to the obtained through-wall radar original echo time-domain data, and then the target area is confirmed on the initial frame radar image. First, a 2D cell-average constant false alarm rate (CA-CFAR) detector is used to detect the radar image to separate the target from the background. The target usually occupies a relatively large area on the radar image. Then, a connected component detection operation is performed on the radar image to mark the connected components with sizes meeting the requirements, and the initial target area can be obtained.
[0054] The maximum value of the pixels in this area is selected as the initial position z0 of the target. Then, with z0 as the center, an image block P containing the target area is extracted from the radar image, and the size and angle information of the initial target area are calculated on P by using the image moment method:
[0055]
[0056]
[0057] Among them, P bina (·) represents a binarized image, θ > 0 indicates that the target rotates clockwise, and vice versa indicates counterclockwise rotation.
[0058] To obtain the scale and angle information of the target area, the initial target area can be directly demarcated on the radar image with a rectangular box. The width N of the rectangular box x = 2a, the height N y = 2b, and the rotation angle is θ.
[0059] According to the width and height of the initial target area, the ideal response map G is set as:
[0060]
[0061] Among them, (x0, y0) is the center position of the response map,
[0062] Step 2: Extract the candidate target area according to the target geometric information of the previous frame, and perform scale and angle transformation to obtain the candidate target area Z:
[0063] 1. At the k-th frame, with the target position z of the previous frame k-1 as the center, extract an image block on the radar image of the current frame Then, according to the rotation angle θ of the target image of the previous frame k-1 for perform rotation:
[0064]
[0065] 2. According to the width (N x ) k-1 and height (N y ) k-1 of the target rectangular box of the previous frame, extract the candidate target area centered on z k-1 on the rotated image slice
[0066] 3. Perform scale transformation on the candidate target area to fix its size to the size of the initial target area:
[0067]
[0068] Step 3: Perform a correlation operation on the candidate target region Z using the filter W to obtain the response map f(Z), where the position corresponding to the maximum response value is the target position z of the current frame. k :
[0069] Based on the candidate target region Z after scale and rotation transformation k and the filter W trained in the previous frame k-1 , calculate the response value:
[0070]
[0071] where F(·) represents the Fourier transform, ⊙ represents the Hadamard product, and α k-1 is the filter parameter of the previous frame, and its calculation formula is:
[0072]
[0073] The position with the maximum response value is the target position z at the current moment k .
[0074] Step 4: Centered on the target position z k , calculate the scale and angle information of the target region, extract the target region on the radar image and perform scale and rotation transformation to obtain the training sample M:
[0075] 1. Centered on the target position z of the current moment k , extract a sub-image containing the target region Then, according to Step 1, calculate the length a of the major axis, k the length b of the minor axis k and the rotation angle θ of the elliptical target region; rotate the sub-image :
[0076]
[0077] 2. Centered on z k , with 2a k as the width and 2b k as the height, extract a rectangular target region on as the training sample;
[0078] 3. Then perform scale transformation on the training sample to fix its size back to the initial target region size:
[0079]
[0080] Step 5: Update the filter W according to the new training sample:
[0081] Update the parameter α of the filter W k as follows:
[0082]
[0083] where M k is the new training sample, G is the ideal response map, * represents the complex conjugate, and the calculation formula of k is:
[0084]
[0085] Remark:
[0086] The shortest regular path length used for quantitatively analyzing the tracking performance in the present invention is defined as:
[0087] dis = S MN
[0088] where M and N are the time series lengths of the tracking trajectory and the reference trajectory, and S is the cumulative cost matrix:
[0089]
[0090] In the above formula, D is the distance matrix:
[0091]
[0092] where T is the required tracking trajectory is the reference trajectory.
[0093] As Figure 1 shown, the simulation uses the gprMax software to construct a through-wall scenario model, and the simulation parameters are as shown in Table 1.
[0094] Table 1 Simulation parameter settings
[0095]
[0096] Figure 2 is the result diagram of processing simulation data by different methods. Among them, (a) is the result diagram of processing simulation data by the mean shift algorithm, (b) is the result diagram of processing simulation data by the shape-adaptive mean shift algorithm, (c) is the result diagram of processing simulation data by the kernel correlation filtering algorithm, (d) is the result diagram of processing simulation data by the rotational kernel correlation filtering algorithm, and (e) is the result diagram of processing simulation data by the method of the present invention. The white dashed line represents the target motion reference trajectory, the red rectangular box or ellipse represents the located target area, and the red dots represent the target tracking positions at each moment. Figure 3It is the position error graph of the simulation data processed by different methods. The mean shift algorithm uses the elliptical template determined by the initial frame to track the target, which easily leads to the failure to match the target subsequently, resulting in a large tracking error; the shape-adaptive mean shift algorithm adds the estimation of the target size and angle, and improves the tracking accuracy by updating the target template in real time, but it is sometimes easily affected by the noise around the target image, resulting in an increase in the tracking error; the kernel correlation filtering algorithm uses the target samples to train the classifier, which can well distinguish the background from the target, but the tracking effect on the target image with rotational changes will be weakened. The rotational kernel correlation filtering algorithm introduces a rotation operation, adjusts the selection angles of the training samples and the candidate regions according to the image changes, and has a good tracking effect on the images with rotational changes. The method proposed in the present invention is adaptable to the target images with scale and rotation angle changes, the tracking trajectory is smooth, and has the minimum tracking error.
[0097] Table 2 Average shortest regularization path length of different methods
[0098]
[0099] The measured scene graph is as Figure 4 shown. In the experiment, a ten-transmitter and ten-receiver MIMO radar is used to collect data, and the transmitted waveform uses a stepped-frequency signal with a frequency band range of 1.9 - 2.412 GHz and a frequency step of 2 MHz.
[0100] Table 2 shows the average shortest regularization path lengths of the method of the present invention and the other four comparison methods for processing simulation data and measured data. It can be seen that the method of the present invention has the shortest average length, indicating that the tracking trajectory of the method of the present invention is the most similar to the reference trajectory and has the highest tracking accuracy. Figure 5 It is the result graph of different methods for processing measured data. Among them, (a) is the result graph of processing measured data using the mean shift algorithm, (b) is the result graph of processing measured data using the shape-adaptive mean shift algorithm, (c) is the result of processing measured data using the kernel correlation filtering algorithm, (d) is the result of processing measured data using the rotational kernel correlation filtering algorithm, and (e) is the result graph of processing measured data using the method of the present invention. The tracking trajectory obtained by the method proposed in the present invention is not only the smoothest, but also the closest to the reference trajectory, and has the best tracking effect.
[0101] The present invention provides a method for tracking through-wall radar targets based on scale adaptive-rotation kernel correlation filtering, which has good tracking performance on continuously changing target images. This method is based on the framework of kernel correlation filtering method, considering the size and angle of the target image, and accurately extracts the target area as the training sample. Then, according to the scale and angle information of the target image, the candidate target area is extracted, and the trained filter is used to achieve the positioning and tracking of the target. By processing and analyzing the experimental data obtained by the through-wall radar, the effectiveness and practicability of this method in tracking moving human targets inside buildings are verified. The results show that the proposed method can effectively track human targets with scale and rotation angle changes in the radar image domain, providing a useful solution for tracking moving human targets inside buildings.
[0102] Of course, the present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can certainly make various corresponding changes and deformations according to the present invention. However, these corresponding changes and deformations should all fall within the protection scope of the appended claims of the present invention.
Claims
1. A through-wall radar target tracking method based on scale-adaptive rotation kernelized correlation filter (SA-RKCF), characterized in that: The following steps are involved: Step 1: extract the initial target area from the initial frame radar image and estimate its scale and angle information, and then set the ideal response graph G; Step 2: Extract the candidate target area according to the target geometry information of the previous frame, and perform scale and angle transformation to obtain the candidate target area Z; Step 3: Use the filter W to perform correlation operation with the candidate target area Z to obtain the response value f(Z), where the position corresponding to the maximum response value is the current frame target position z; Step 4: Taking the target position z as the center, calculate the scale and angle information of the target area, extract the target area on the radar image, and perform scale and rotation transformation to obtain the training sample M; Step 5: Update the filter W based on the new training samples.
2. A through-wall radar target tracking method based on scale adaptive-rotation kernel correlation filtering as claimed in claim 1, characterized in that: In step 1, the calculation method of the size and angle information of the initial target area is: The maximum value of the pixel in the initial target area is selected as the initial target position z0, and then an image block P including the target area is extracted from the radar image with z0 as the center. The size and angle information of the initial target area are calculated on P using the image moment method: Among them, P bina (·) represents the binary image, θ>0 means the target rotates clockwise, otherwise it rotates counterclockwise.
3. The through-wall radar target tracking method based on scale adaptive-rotation kernel correlation filtering according to claim 1, characterized in that: In step 1, according to the initial target area width and height, the ideal response graph G is set as: Where (x0, y0) is the center position of the response graph, σ is the corresponding variance on the x-axis and y-axis, 4. The through-wall radar target tracking method based on scale adaptive-rotation kernel correlation filtering according to claim 1, characterized in that: In step 2, the method for obtaining the candidate target area Z is:
1. In the kth frame, the target position z of the previous frame k-1 As the center, extract an image block on the current frame radar image Then rotate the angle θ according to the target image of the previous frame k-1 right To rotate:
2. According to the target rectangle width (N x ) k-1 and height (N y ) k-1 , after rotation of the image slice z k-1 Extract candidate target areas for the center 3. Transform the scale of the candidate target area and fix its size back to the initial target area size:
5. The through-wall radar target tracking method based on scale adaptive-rotation kernel correlation filtering according to claim 1, characterized in that: Furthermore, in step 3, the calculation method of the response value f(Z) is: According to the candidate target area Z after scale and rotation transformation k The filter W obtained by training the previous frame k-1 , calculate the response value: Where F(·) represents Fourier transform, ⊙ represents Hadamard product, α k-1 is the filter parameter of the previous frame, The calculation formula is:
6. The through-wall radar target tracking method based on scale adaptive-rotation kernel correlation filtering according to claim 1, characterized in that: In step 3, the position with the largest response value is the current target position z k .
7. The through-wall radar target tracking method based on scale adaptive-rotation kernel correlation filtering as claimed in claim 1, characterized in that: In step 4, the method for obtaining the training sample M is:
1. Take the current target position z k As the center, extract a sub-image containing the target area Then follow step 1 to calculate the length a of the major semi-axis of the elliptical target area. k , the length of the minor axis b k and rotation angle θ; the sub-image To rotate:
2. With z k Centered, 2a k is the width, 2b k For height, Extract the rectangular target area As training samples; 3. Then rescale the training samples and fix their size to the initial target area size:
8. The through-wall radar target tracking method based on scale adaptive-rotation kernel correlation filtering as claimed in claim 1, characterized in that: In step 5, the update method of the filter W is: The parameter α of the filter W k To update: Among them, M k is a new training sample, G is the ideal response graph, F(·) represents Fourier transform, * represents complex conjugate, and the calculation formula of k is: