A through-the-wall radar moving human body target tracking method based on rotation kernel correlation filtering

Abstract

In order to solve the problem that the existing method cannot effectively solve the difficulty in tracking the radar image of the target with rotation change, a through-wall radar moving human target tracking method based on rotation kernel correlation filtering is disclosed.The method comprises the following steps: firstly, a candidate target region is extracted on the radar image, and the region image is rotated after the internal target ellipse angle is estimated; then, a classifier is used for correlation operation with the candidate region, so as to locate the target; finally, new training samples are extracted to update the classifier.Compared with the existing through-wall radar image domain tracking method, the proposed method has better tracking accuracy and good real-time performance.Numerical simulation and measured results prove the effectiveness of the proposed algorithm.

Description

Technical Field

[0001] This invention belongs to the field of through-wall radar, and particularly relates to a method for tracking moving human targets using through-wall radar based on rotating kernel correlation filtering. Background Technology

[0002] With urbanization, urban warfare will become an important form of future warfare, and the obstruction of buildings makes it difficult to comprehensively grasp battlefield information. Through-the-wall radar (TWR) utilizes the penetrating power of electromagnetic waves to detect indoor targets and can acquire information about human targets behind obstacles. However, the fixed observation angle of TWR and the complex scattering characteristics of human targets cause the radar image of the target to change in scale and angle with the position and movement of the human body, making target tracking difficult. Therefore, there is an urgent need to study human target image domain tracking algorithms that can adapt to changes in the target radar image, providing technical support for real-time monitoring of enemy movements.

[0003] Kalman filtering, which combines measurements with a system model to predict and correct system state values, is widely used in through-wall radar target tracking. However, this algorithm requires accurate motion modeling of the target; otherwise, tracking accuracy will decrease. Mean shift algorithm can track targets using radar image features without requiring motion modeling. However, this algorithm is susceptible to image noise and has a long computation time. Kernel correlation filtering algorithm uses target image features to train a classifier to distinguish between background and target, achieving target localization and tracking. However, this algorithm struggles to adapt to target images that have undergone rotational changes, resulting in poor tracking performance. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a through-wall radar moving human target tracking algorithm based on rotating kernel correlation filtering. This algorithm can accurately track targets with changing angles on radar images and also exhibits good real-time performance.

[0005] The technical solution of this invention is as follows:

[0006] A method for tracking moving human targets using through-wall radar based on rotating kernel correlation filtering includes the following steps:

[0007] Step 1: Sampling of the target area;

[0008] Imaging is performed based on the obtained raw echo time-domain data from the through-wall radar. Then, the desired region samples are extracted from the radar image. In the initial frame, the target region size is determined using a biaxial projection method, with the pixel projection values ​​of the target region on the X and Y axes as follows:

[0009]

[0010] Where N x and N y These are the width and height of the target area, respectively.

[0011] Set a threshold value to determine T. x and T y If the first and last elements exceed a threshold, the target region is expanded in that direction. This process is repeated until the projection value is less than the threshold. The target region is then defined by the final width N. x and height N y To be determined.

[0012] In subsequent frames, using the target position from the previous frame as the center, and based on the width N... x and height N y Extract candidate target region samples Z from radar images.

[0013] Step 2: Rotate the candidate target region;

[0014] In through-wall radar images, target images typically exhibit an elliptical shape. The rotation angle of this ellipse can be estimated using the image moment method.

[0015]

[0016]

[0017] Among them, I bina (·) represents a binary image, θ>0 indicates that the target rotates clockwise, and vice versa.

[0018] Rotate the candidate target region sample to a horizontal position based on the estimated angle:

[0019] Z rotation =imrotate(Z,θ)

[0020] Step 3: Target location estimation;

[0021] Based on the rotated candidate target region sample Z rotation Compared with the classifier W trained in the previous frame k-1 Calculate the response value:

[0022]

[0023] Where F(·) represents the Fourier transform, ⊙ represents the Hadamard product, and α is the classifier parameter. The calculation formula is:

[0024]

[0025] The position with the largest response value is the target location y at the current moment.k .

[0026] Step 4: Update the classifier;

[0027] Centered on the current target location, N x and N y The target region M is extracted based on its width and height, and used as a new training sample. The target angle at the current moment is estimated using the angle estimation method from step 2, and the training sample is updated by rotation. Then, based on the latest training sample M... rotation Update the classifier:

[0028]

[0029] Where G is the ideal response diagram, * denotes complex conjugate, and k is calculated using the following formula:

[0030]

[0031] In this invention, the shortest regularization path length used for quantitative analysis is defined as:

[0032] dis = S MN

[0033] Where M and N are the time series lengths of the tracking trajectory and the reference trajectory, respectively, and S is the cumulative cost matrix:

[0034]

[0035] In the above formula, D is the distance matrix:

[0036]

[0037] Where y is the desired tracking trajectory. For reference trajectory.

[0038] Beneficial effects:

[0039] This invention provides a method for tracking moving human targets using through-wall radar based on rotating kernel correlation filtering. By improving upon the kernel correlation filtering algorithm framework, the algorithm achieves good real-time performance. Simulation and experimental results show that, compared with other methods, this invention has better tracking accuracy for rotating targets, making it an effective method for tracking moving human targets using through-wall radar. Attached Figure Description

[0040] Figure 1 This is a scene diagram used in the simulation of the method of this invention;

[0041] Figure 2These are the results of processing simulation data using different methods. Among them, (a) is the result of processing simulation data using the mean shift algorithm, (b) is the result of processing simulation data using the shape adaptive mean shift algorithm, (c) is the result of processing simulation data using the kernel correlation filter algorithm, and (d) is the result of processing simulation data using the method of this invention.

[0042] Figure 3 These are position error diagrams for processing simulation data using different methods;

[0043] Figure 4 These are two sets of scene diagrams and radar array distribution diagrams obtained from actual measurements using the method of this invention;

[0044] Figure 5 and Figure 6 The figures show the results of processing two sets of measured data using different methods. (a) shows the result of processing the measured data using the mean shift algorithm, (b) shows the result of processing the measured data using the shape adaptive mean shift algorithm, (c) shows the result of processing the measured data using the kernel correlation filter algorithm, and (d) shows the result of processing the measured data using the method of this invention.

[0045] Specific implementation process

[0046] The present invention will be further described below with reference to the embodiments.

[0047] Step 1: Sampling of the target area;

[0048] Imaging is performed based on the obtained raw echo time-domain data from the through-wall radar. Then, the desired region samples are extracted from the radar image. In the initial frame, the target region size is determined using a biaxial projection method, with the pixel projection values ​​of the target region on the X and Y axes as follows:

[0049]

[0050] Where N x and N y These are the width and height of the target area, respectively.

[0051] Set a threshold value to determine T. x and T y If the first and last elements exceed a threshold, the target region is expanded in that direction. This process is repeated until the projection value is less than the threshold. The target region is then defined by the final width N. x and height N y To be determined.

[0052] In subsequent frames, using the target position from the previous frame as the center, and based on the width N... x and height N y Extract candidate target region samples Z from radar images.

[0053] Step 2: Rotate the candidate target region;

[0054] In through-wall radar images, target images typically exhibit an elliptical shape. The rotation angle of this ellipse can be estimated using the image moment method.

[0055]

[0056] Among them, I bina (·) represents a binary image, θ>0 indicates that the target rotates clockwise, and vice versa.

[0057] Rotate the candidate target region sample to a horizontal position based on the estimated angle:

[0058] Z rotation =imrotate(Z,θ)

[0059] Step 3: Target location estimation;

[0060] Based on the rotated candidate target region sample Z rotation Compared with the classifier W trained in the previous frame k-1 Calculate the response value:

[0061]

[0062] Where F(·) represents the Fourier transform, ⊙ represents the Hadamard product, and α is the classifier parameter. The calculation formula is:

[0063]

[0064] The position with the largest response value is the target location y at the current moment. k .

[0065] Step 4: Update the classifier;

[0066] Centered on the current target location, N x and N y The target region M is extracted based on its width and height, and used as a new training sample. The target angle at the current moment is estimated using the angle estimation method from step 2, and the training sample is updated by rotation. Then, based on the latest training sample M... rotation Update the classifier:

[0067]

[0068] Where G is the ideal response diagram, * denotes complex conjugate, and k is calculated using the following formula:

[0069]

[0070] In this invention, the shortest regularization path length used for quantitative analysis is defined as:

[0071] dis = S MN

[0072] Where M and N are the time series lengths of the tracking trajectory and the reference trajectory, respectively, and S is the cumulative cost matrix:

[0073]

[0074] In the above formula, D is the distance matrix:

[0075]

[0076] Where y is the desired tracking trajectory. For reference trajectory.

[0077] like Figure 1 As shown in Table 1, the simulation was performed using gprMax software to construct a wall-penetrating scenario model.

[0078] Table 1 Simulation parameter settings

[0079]

[0080] Figure 2 These are the results of processing simulation data using different methods. (a) shows the result using the mean-shift algorithm, (b) shows the result using the shape-adaptive mean-shift algorithm, (c) shows the result using the kernel correlation filter algorithm, and (d) shows the result using the method of this invention. The white dashed line represents the target motion reference trajectory, the red rectangle or ellipse represents the located target area, and the red dots represent the target tracking position at each time step. Figure 3 This diagram illustrates the positional error of simulation data processed by different methods. The mean-shift algorithm uses an elliptical template determined in the initial frame to track the target, which can easily lead to subsequent mismatches and significant tracking errors. The shape-adaptive mean-shift algorithm adds target size and angle estimation and improves tracking accuracy by updating the target template in real time; however, it is sometimes susceptible to noise around the target image, increasing tracking error. The kernel correlation filter algorithm uses target samples to train a classifier, which can effectively distinguish between the background and the target, but its tracking performance weakens for target images undergoing rotation. The method proposed in this invention is adaptable to rotating target images, provides a smooth tracking trajectory, and achieves the best tracking accuracy.

[0081] Table 2. Average shortest regularization path length for different methods

[0082]

[0083] Actual test scenario diagram as shown Figure 4 As shown, the experiment used a MIMO radar with ten transmitters and ten receivers to collect data. The transmitted waveform used a stepped frequency signal with a frequency band range of 1.9-2.412GHz and a frequency step of 2MHz.

[0084] Table 2 shows the average shortest regularized path lengths of the method of this invention and the other three comparative methods when processing measured data. It can be seen that the method of this invention has the shortest average length, indicating that the tracking trajectory of this invention is most similar to the reference trajectory, and thus has the highest tracking accuracy. Figure 5 and Figure 6 These are the results of processing two sets of measured data using different methods. (a) shows the result using the mean-shift algorithm, (b) shows the result using the shape-adaptive mean-shift algorithm, (c) shows the result using the kernel correlation filter algorithm, and (d) shows the result using the method of this invention. The tracking trajectory obtained by the method proposed in this invention is not only the smoothest but also closest to the reference trajectory, resulting in the best tracking performance.

[0085] Table 3 Comparison of processing times for different methods

[0086]

[0087] Table 3 compares the processing time of the method of this invention with the other three comparative methods. Each scenario was repeated 20 times. In the simulated scenario, the radar image is relatively clean, and the mean-drift algorithm requires fewer iterations to find the target region, resulting in less tracking time. The scale-adaptive mean-drift algorithm adds a shape estimation operation, making its tracking time longer than the mean-drift algorithm. The kernel correlation filter algorithm directly locates the target region using a trained classifier without iterative calculations, effectively reducing tracking time. The method of this invention is an improvement on the kernel correlation filter algorithm framework, adding angle estimation and rotation operations. Although it takes longer than the kernel correlation filter algorithm, it still exhibits excellent real-time performance compared to mean-drift algorithms.

[0088] In experimental scenarios, radar images often contain numerous ghosting patterns, increasing the number of iterations required for mean-drift algorithms to find the target region and significantly increasing tracking time. While the scale-adaptive mean-drift algorithm adds a shape estimation operation, its real-time updated target template reduces the number of iterations needed to converge to the target region, resulting in a shorter overall tracking time compared to the mean-drift algorithm. Both the kernel correlation filtering algorithm and the method described in this invention are less affected by image quality and can still process over a hundred radar images per second, demonstrating excellent real-time performance in practical tracking scenarios.

[0089] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A method for tracking moving human body targets by through-the-wall radar based on rotated kernel correlation filtering, characterized in that, Includes the following steps: Step 1: Target region sampling: Extract candidate target regions Z from the radar image; Step 2: Candidate target region rotation: angle estimation for the target ellipse within the candidate target region and rotated to horizontal to get The specific method for angle estimation for the target ellipse within the candidate target region and rotation is as follows: ； ； ； ； ； wherein denotes a binarized image, denotes a clockwise rotation of the object, and vice versa; Step 3: Target position estimation: using the classifier W and the rotated candidate target region Perform correlation operation to obtain response Where the position corresponding to the maximum response is the target positioning position y; according to the rotated candidate region And the classifier obtained in the previous frame Calculate the response value to obtain the target positioning value y k : ； ； ； wherein, denotes a Fourier transform, denotes a Hadamard product, is a classifier parameter; Step 4: Update the classifier: extract target region training samples M on the radar image centered on the target location, and perform angle estimation on the inner target ellipse and rotate the training samples to horizontal, update the classifier.

2. The method for tracking moving human targets using through-wall radar based on rotating kernel correlation filtering as described in claim 1, characterized in that, In the initial frame, determine the target region width N x and height N y In the subsequent frame, extract the candidate target region centered on the target positioning position of the previous frame Z .

3. The method for tracking moving human targets using through-wall radar based on rotating kernel correlation filtering as described in claim 1, characterized in that... , the target positioning position at the current time y k extracting target region training samples on the radar image M angle estimation on the internal target ellipse and rotating the training samples to horizontal updating the classifier: ； ； in, G For the ideal response diagram, Indicates complex conjugation.

Images