A polsar vehicle target detection method based on fine polarization decomposition

Abstract

The application discloses the technical field of polarimetric synthetic aperture radar vehicle target detection and discloses a PolSAR vehicle target detection method based on fine polarization decomposition. The PolSAR vehicle target detection method based on fine polarization decomposition comprises the following steps: a rotation dihedral scattering model is constructed, a fine five-component polarization decomposition method is proposed, and fine polarization decomposition is carried out on PolSAR vehicle target scattering; based on a stable scattering power composite feature, a scattering power composite feature detector is constructed, and effective PolSAR vehicle target detection is realized. The PolSAR vehicle target detection method based on fine polarization decomposition starts from fine characterization of target scattering behavior, proposes a fine five-component polarization decomposition method, and constructs a stable scattering power composite feature to distinguish vehicle targets from background clutter and artificial clutter interference, thereby achieving simple and effective detection of vehicle targets.

Description

Technical Field

[0001] This invention relates to the field of polarimetric synthetic aperture radar (PolSAR) vehicle target detection technology, specifically to a PolSAR vehicle target detection method based on fine polarimetric decomposition. Background Technology

[0002] Polarimetric synthetic aperture radar (PolSAR) possesses all-weather, day-and-night ground imaging capabilities and can simultaneously acquire data from different polarization channels, making it an indispensable sensor in the field of remote sensing. PolSAR systems acquire data containing rich polarization information about targets, promoting the development of target scattering mechanism interpretation technology and playing a crucial role in the detection of man-made targets such as buildings, ships, and vehicles. Among these, PolSAR vehicle target detection is particularly critical for military and civilian applications such as battlefield environmental reconnaissance and surveillance, and road traffic monitoring. Currently, there is limited research on PolSAR vehicle target detection in the literature. This is partly because the imaging scenes of vehicle targets are generally complex, including grasslands, trees, buildings, and roads, making effective detection difficult based solely on intensity information. Furthermore, the availability of publicly available PolSAR vehicle datasets is limited, hindering related research. Generally, targets in simple scenes are spatially dispersed, with uniform and low-intensity background clutter and minimal man-made clutter interference (e.g., ships on calm sea surfaces), making their detection relatively easy. In complex scenarios, targets are spatially concentrated, leading to mutual interference among multiple targets. Furthermore, background clutter is heterogeneous, high-intensity, and heavily influenced by artificial clutter, making detection difficult and rendering traditional target detection methods based on scattering intensity inapplicable. Additionally, when a vehicle target's orientation is not parallel to the radar's flight direction (referred to as a variable-azimuth vehicle target), its backscattering produces a strong cross-polarization response, posing a challenge to PolSAR data interpretation. Therefore, to achieve vehicle target detection in complex scenarios, it is necessary to accurately characterize its scattering behavior and construct robust and effective polarization features to distinguish it from background clutter and artificial clutter interference. Summary of the Invention

[0003] The purpose of this invention is to provide a PolSAR vehicle target detection method based on fine polarization decomposition, which solves the technical problem of difficulty in detecting PolSAR vehicle targets in complex scenes in the prior art. Starting from the fine characterization of target scattering behavior, a fine five-component polarization decomposition method is proposed, and a robust scattering power composite feature is constructed to distinguish vehicle targets from background clutter and artificial clutter interference, so as to achieve simple and effective detection of vehicle targets.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows:

[0005] This invention proposes a PolSAR vehicle target detection method based on fine polarization decomposition, comprising the following steps:

[0006] S1. Construct a rotating dihedral scattering model and propose a fine five-component polarization decomposition method to perform fine polarization decomposition on the scattering of PolSAR vehicle targets;

[0007] S2. Based on robust scattering power recombination features, a scattering power recombination feature detector is constructed to achieve effective detection of PolSAR vehicle targets.

[0008] As one aspect of the PolSAR vehicle target detection method based on fine polarization decomposition of the present invention, S1 includes the following steps:

[0009] S11. For vehicle target detection tasks, a rotating dihedral scattering model RDSM was constructed based on GRDSM;

[0010] S12. A refined five-component polarization decomposition method is proposed by combining the rotating dihedral angle scattering model RDSM.

[0011] As one aspect of the PolSAR vehicle target detection method based on fine polarization decomposition of the present invention, S11 includes the following steps:

[0012] S111. A reasonable model for scattering of rotating dihedral structures is proposed. Based on the cross-scattering model CSM, a generalized rotating dihedral scattering model GRDSM is proposed, with its coherence matrix [T]. GR The expression is as follows:

[0013]

[0014] Among them, X 22 +X 33 =1, 0≤X 22 ,X 33 ≤1 and X 22 ≤X 33 ;

[0015] S112. Based on the generalized rotating dihedral scattering model GRDSM, a rotating dihedral scattering model RDSM is constructed, with its coherence matrix [T]. RDSM The expression is as follows:

[0016]

[0017] D in the above formula RDS The expression is derived from the eigenvalues ​​of the coherence matrix and is as follows:

[0018]

[0019] Where SPAN represents the total scattered power, λ i (i = 1, 2, 3) represent the eigenvalues ​​of the coherence matrix.

[0020] As one aspect of the PolSAR vehicle target detection method based on fine polarization decomposition of the present invention, the coherence matrix [T] of the cross-scattering model CSM in S111 is... cross The expression is as follows:

[0021]

[0022] Where, θ dom This indicates the dominant polarization azimuth angle.

[0023] As one aspect of the PolSAR vehicle target detection method based on fine polarization decomposition of the present invention, S12 includes the following steps:

[0024] S121. The expression for the refined five-component polarization decomposition method is as follows:

[0025] <[T]> = f S [T] S +f D [T] D +f V [T] V +f H [T] H +f R [T] RDSM

[0026] Where <[T]> represents the measurement coherence matrix, [T] S [T] D [T] V [T] H and [T] RDSM These correspond to surface scattering, even-order scattering, volume scattering, helical scattering, and rotating dihedral scattering models, respectively, f S f D f V f H and f R Let [T] represent the corresponding expansion coefficients. S [T] D [T] V and [T] H The expression is as follows:

[0027]

[0028]

[0029] S122. Expand the measurement coherence matrix <[T]> element by element to obtain the following system of equations:

[0030]

[0031]

[0032] f S β * +f D α=T 12

[0033]

[0034]

[0035] S123. Using branch condition T 11 -T 22 +f H / 2 Fix some model parameters of the equation system, when T 11 -T 22 +f H When / 2>0, the dominant scattering mechanism of the remaining matrix is ​​considered to be surface scattering, let f D =0 (α=0), otherwise, even-order scattering is considered the dominant scattering mechanism of the residual matrix, let f S =0 (β=0);

[0036] S124. In non-rotational dihedral regions or other natural regions, the equations in the system of equations The values ​​are ignored during the calculation, and the same applies to the dihedral region of rotation. The solution for all model parameters is as follows:

[0037] T 11 -T 22 +f H / 2>0:

[0038]

[0039]

[0040]

[0041] T 11 -T 22 +f H / 2≤0:

[0042]

[0043]

[0044] f V =2(2T) 22 -2f D -f H ),

[0045]

[0046] Therefore, the power of each scattering component is as follows:

[0047] P S =f S (1+|β| 2 ),P D =f D (1+|α| 2 ),P H =f H ,P R =f R ,

[0048] P V =SPAN-P S -P D -P H -P R .

[0049] As one aspect of the PolSAR vehicle target detection method based on fine polarization decomposition of the present invention, wherein S2 includes the following steps:

[0050] S21. A simple and physically clear composite feature of scattering power was constructed.

[0051] S22. A composite feature detector for scattering power was constructed to achieve effective detection of PolSAR vehicle targets.

[0052] As one aspect of the PolSAR vehicle target detection method based on fine polarization decomposition of the present invention, wherein S21 includes the following steps:

[0053] S211. For azimuth-parallel vehicle targets, a composite scattering power feature is proposed. para-azi Its expression is as follows:

[0054] Feature para-azi =P S ×P D ×P V ;

[0055] S212. For vehicles with varying azimuth, a composite scattering power feature is proposed.vari-azi Its expression is as follows:

[0056] Feature vari-azi =(P R +P H )×P V ;

[0057] Used to construct composite features para-azi and Feature vari-azi The power of each scattering component has been normalized, and the expression is as follows:

[0058]

[0059] Where i = S, D, V, H, R, P i The value range is [0,1];

[0060] S213. Combining Feature para-azi and Feature vari-azi The feature of scattering power recombination was proposed. vehicle Its expression is as follows:

[0061] Feature vehicle =Feature para-azi +Feature vari-azi

[0062] =P S ×P D ×P V +(P R +P H )×P V .

[0063] As one aspect of the PolSAR vehicle target detection method based on fine polarization decomposition of the present invention, wherein S22 includes the following steps:

[0064] S221. Feature Based on Scattering Power Recombination Characteristics vehicle Construct a scattering power composite feature detector Det vehicle As shown in the following formula:

[0065] Det vehicle =Density(Feature) vehicle )

[0066] Where Density(X) represents the density of X, which is the mean value of X in the N×N neighborhood of the center pixel, and the size of the neighborhood is set to 3×3;

[0067] S222. Set the detection threshold to THvehicle The PolSAR vehicle target detection results are as follows:

[0068]

[0069] As one aspect of the PolSAR vehicle target detection method based on fine polarization decomposition of the present invention, wherein the detection threshold TH in S222 vehicle The method for determining it includes the following steps:

[0070] S2221. Select a typical ground background region as training data and calculate Det within that region. vehicle The value;

[0071] S2222. Obtain the Det vehicle The values ​​are sorted in ascending order, and the largest values, which account for 0.1% to 0.5% of the total number of pixels in the training data, are removed. The remaining pixels' Det values ​​are then... vehicle The maximum value is determined to be TH vehicle .

[0072] By adopting the above technical solution, the present invention has the following advantages:

[0073] This invention provides a PolSAR vehicle target detection method based on fine polarization decomposition. For vehicle target detection in complex scenarios, a rotating dihedral scattering model is introduced, fully considering the rapid transformation between co-polarization and cross-polarization responses caused by changes in the dihedral structure's orientation. Based on this, a fine five-component polarization decomposition method is proposed, enabling quantitative characterization of vehicle target scattering behavior. Furthermore, by analyzing the scattering differences between the vehicle target and the ground background, this invention constructs a simple and physically meaningful composite scattering power feature detector, enhancing the scattering difference between the vehicle target and the ground background, and achieving robust and effective vehicle target detection. Attached Figure Description

[0074] Figure 1 Analysis of vehicle target scattering composition for the present invention;

[0075] Figure 2 UAV-borne SAR data 1 and data 2 used for vehicle target detection in this invention;

[0076] Figure 3 The result of fine five-component polarization decomposition of UAV-borne SAR data 1 of the present invention;

[0077] Figure 4 This is the result of the fine five-component polarization decomposition of the UAV-borne SAR data 2 of the present invention;

[0078] Figure 5Amplitude plots of different detectors on UAV-borne SAR data 1;

[0079] Figure 6 Amplitude plots of different detectors on UAV-borne SAR data 2;

[0080] Figure 7 The TCR variation of the detector of this invention and other detectors at various vehicle targets;

[0081] Figure 8 ROC curves of the detector of this invention and other detectors;

[0082] Figure 9 The detection results of the method of the present invention and other methods on UAV-borne SAR data 1;

[0083] Figure 10 The results of the present invention and other methods on UAV-borne SAR data 2 are shown. Detailed Implementation

[0084] The technical solution of the present invention will be specifically described below with reference to the accompanying drawings. It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0085] PolSAR systems acquire rich polarization information about targets by transmitting and receiving orthogonal electromagnetic waves. The measured PolSAR data can be represented as a 2×2 scattering matrix S.

[0086]

[0087] In this context, the subscripts H and V represent horizontal and vertical polarization, respectively. HV This represents the complex scattering coefficient for vertically polarized transmission and horizontally polarized reception; the definitions of the remaining terms are similar. In the case of monostatic backscattering, S generally satisfies the reciprocity theory, i.e., S HV =S VH .

[0088] The target vector k based on the Pauli basis p It can be represented as

[0089]

[0090] Where the superscript T denotes transpose. From the target vector k p The polarization coherence matrix T can be obtained.

[0091]

[0092] The superscripts * and H denote conjugate and conjugate transpose, respectively, and <·> denote spatial average. The coherence matrix T is a positive semi-definite Hermitian matrix with real diagonal elements and complex off-diagonal elements, and a total of 9 degrees of freedom.

[0093] Dihedral structures are commonly found in man-made target scenarios such as vehicles, ships, and buildings. The even-order scattering model commonly used in polarization decomposition is suitable for describing the scattering produced by dihedral structures parallel to the radar flight direction (called non-rotating dihedrals). When the azimuth angle of the dihedral structure changes in the imaging scene (called a rotating dihedral), its backscattering can generate strong cross-polarization energy, rendering the traditional even-order scattering model inapplicable. Therefore, to accurately interpret the scattering mechanism of rotating man-made targets, it is necessary to reasonably model the scattering of rotating dihedral structures.

[0094] This invention provides a PolSAR vehicle target detection method based on fine polarization decomposition, comprising the following steps:

[0095] S1. Construct a rotating dihedral scattering model and propose a fine five-component polarization decomposition method to perform fine polarization decomposition on the scattering of PolSAR vehicle targets;

[0096] S1 includes the following steps:

[0097] S11. For vehicle target detection tasks, a rotating dihedral scattering model RDSM was constructed based on GRDSM;

[0098] S11 includes the following steps:

[0099] S111. A reasonable modeling of scattering of rotating dihedral structures is proposed, and a generalized rotating dihedral scattering model GRDSM is proposed based on the cross scattering model CSM.

[0100] Early on, Xiang Deliang et al. modeled cross-polarization scattering generated by rotating dihedral structures, proposing the cross scattering model (CSM). The coherence matrix of this model is [T]. cross It can be represented as

[0101]

[0102] Where, θ dom This indicates the dominant polarization azimuth. It can be observed that only T in the CSM...22 and T 33 The non-zero elements in this matrix form help separate the component from the total cross-polarization energy, improving the overestimation of volume scattering at rotating artificial targets. However, the CSM still has some limitations. First, T in the CSM... 22 and T 33 The sizes are almost equal, with a maximum difference of only 1 / 15. However, Guinvarc'h et al. pointed out that the cross-polarization energy generated by rotating the dihedral angle is much stronger than the same-polarization energy. Therefore, in CSM, T... 22 and T 33 The value of this parameter does not match the actual situation. Secondly, the dominant polarization azimuth angle θ is introduced in CSM. dom This results in the model being significantly affected by azimuth angle estimation, leading to poor robustness.

[0103] To provide a reasonable model for the scattering of rotating dihedral structures, a generalized rotating dihedral scattering model (GRDSM) is proposed based on the CSM, with its coherence matrix [T]. GR The expression is as follows:

[0104]

[0105] Among them, X 22 +X 33 =1, 0≤X 22 ,X 33 ≤1 and X 22 ≤X 33 Considering that the cross-polarization response generated by backscattering from a rotating dihedral structure is generally stronger than the same-polarization response, this invention limits X... 22 ≤X 33 GRDSM provides a standard, unified form for the rotating dihedral scattering model, whose model element X... 22 and X 33 The value of should satisfy the above constraints, and should make the ratio of the cross-polarization component to the same polarization component close to the actual situation.

[0106] S112. For vehicle target detection tasks, a rotated dihedral scattering model (RDSM) with the following parameter expression is constructed based on GRDSM.

[0107]

[0108] D in equation (6) RDS The expression is derived from the eigenvalues ​​of the coherence matrix.

[0109]

[0110] Where SPAN represents the total scattered power, λ i (i = 1, 2, 3) represent the eigenvalues ​​of the coherence matrix.

[0111] D RDS It is related to the target's depolarization characteristics, scattering randomness, and polarization asymmetry, and can well describe the scattering characteristics of a rotating dihedral. Generally speaking, the greater the degree of azimuth rotation of the rotating dihedral, the better the scattering characteristics of D. RDS The larger the value of D, the better. However, in areas without rotational dihedral angles or natural features, D... RDS The value of is relatively small. As can be seen from equation (7), D RDS The range of values ​​for is related to the eigenvalues, and theoretically is [0, ∞). For D RDS Using the sigmoid function transformation, the range of R is [1 / 2, 1). Therefore, the model elements of RDSM always satisfy... It conforms to the general constraints given by GRDSM. Compared with CSM, the proposed RDSM can more accurately describe rotating dihedral scattering, and it has the following characteristics:

[0112] (1) RDSM can separate the component from the rotating dihedral structure from the total cross-polarization energy, thereby improving the volume scattering overestimation at rotating artificial targets;

[0113] (2) For the rotating dihedral structure, the same polarization component in the RDSM is always smaller than the cross polarization component, which is more consistent with the actual scattering behavior of the rotating dihedral structure, and thus can enhance the rotating dihedral scattering.

[0114] (3) The model parameters of RDSM are derived from the eigenvalues ​​of the coherence matrix, so it has rotation invariance and is not affected by the accuracy of azimuth estimation.

[0115] S12. A refined five-component polarization decomposition method is proposed by combining the rotating dihedral angle scattering model RDSM.

[0116] S12 includes the following specific steps:

[0117] S121. The expression for the fine five-component polarization decomposition method can be expressed as follows:

[0118] <[T]>=f S [T] S +f D [T] D +f V [T] V +f H [T]H +f R [T] RDSM (8)

[0119] Where <[T]> represents the measurement coherence matrix, and [T]... S [T] D [T] V [T] H and [T] RDSM These correspond to surface scattering, even-order scattering, volume scattering, helical scattering, and rotating dihedral scattering models, respectively, f S f D f V f H and f R These represent the corresponding expansion coefficients. [T] S [T] D [T] V and [T] H The expression is:

[0120]

[0121]

[0122] S122. Expand the measurement coherence matrix <[T]> element by element, that is, expand the coherence matrix of equation (8) element by element, and you can get the equation set shown in equation (11).

[0123]

[0124] S123. The system of equations shown in equation (11) has 7 unknowns. Therefore, the solution of the model parameters is an underdetermined problem. It is necessary to use branch conditions to fix some of the model parameters. The branch condition used is T. 11 -T 22 +f H / 2.

[0125] When T 11 -T 22 +f H When / 2>0, the dominant scattering mechanism of the remaining matrix is ​​considered to be surface scattering, let f D =0 (α=0). Otherwise, we assume that even-order scattering is the dominant scattering mechanism of the residual matrix, and let f S =0 (β=0). At this point, the number of remaining unknown parameters equals the number of equations, and equation (11) has a compact characteristic. However, directly solving the analytical solution of the model is still very difficult.

[0126] S124. In fact, in non-rotational dihedral regions or other natural regions, f R The value of should be very small, therefore the value in equation (11) is... The value is very small and can be ignored. However, for the dihedral region of rotation, D... RDS The values ​​are very large, so the element 1-R in RDSM is close to zero and can also be ignored. After that, the solution for all model parameters can be obtained.

[0127]

[0128]

[0129] Therefore, the power of each scattering component is

[0130] P S =f S (1+|β| 2 ),P D =f D (1+|α| 2 ),P H =f H ,P R =f R (14)

[0131] P V =SPAN-P S -P D -P H -P R (15)

[0132] S2. Based on robust scattering power recombination features, a scattering power recombination feature detector is constructed to achieve effective detection of PolSAR vehicle targets.

[0133] S2 includes the following specific steps:

[0134] S21. The ground background where vehicle targets are located is generally complex (roads, grass, trees, buildings, etc.), which poses a challenge to PolSAR vehicle target detection. Based on the refined five-component polarization decomposition method, the scattering behavior of vehicle targets can be quantitatively characterized. By analyzing the differences in scattering characteristics between vehicle targets and the ground background, and combining the results of refined five-component polarization decomposition, this invention constructs a simple and physically meaningful composite scattering power feature for vehicle target detection.

[0135] S21 includes the following steps:

[0136] S211. Figure 1 A schematic diagram of the scattering mechanism of vehicle targets is given. Among them, Figure 1In the diagram, (a) represents a vehicle target with parallel azimuth; (b) represents a vehicle target with variable azimuth. A represents surface scattering, B represents even-order scattering, C represents volume scattering (multiple interactive scattering), D represents helical scattering, and E represents dihedral scattering. For a vehicle target with parallel azimuth (its orientation is parallel or perpendicular to the radar flight direction), such as... Figure 1 As shown in (a), the backscattering mechanism mainly includes even-order scattering generated by the dihedral structure formed by the vehicle body and the ground, surface scattering generated by parts such as the roof, and multiple interactive scattering generated by the interaction between the vehicle body and the ground background. To highlight azimuthally parallel vehicle targets in the ground background, this invention proposes a scattering power composite feature. para-azi Its expression is:

[0137] Feature para-azi =P S ×P D ×P V (16)

[0138] From equation (16), it can be found that Feature para-azi It emphasizes azimuth-parallel vehicle targets where surface scattering, even-order scattering, and multiple interactive scattering are all significant.

[0139] For ground backgrounds such as roads, grass, and trees, since they only have strong surface scattering or volume scattering power, the resulting features... para-azi Very small. The power of each scattering component used to construct the composite feature has been normalized, i.e.

[0140]

[0141] Where i = S, D, V, H, R. After normalization, P i The value range is [0,1].

[0142] S212. For vehicle targets with variable azimuth (the vehicle target's orientation is not parallel to the radar's flight direction), such as Figure 1 As shown in (b), its backscattering can generate strong cross-polarization energy. Its scattering mechanisms mainly include rotating dihedral scattering generated by the rotating dihedral structure formed by the vehicle body and the ground, surface scattering from parts such as the roof, multiple interactive scattering generated by the interaction between the vehicle body and the ground background, and spiral scattering. To highlight the variable-orientation vehicle target in the ground background, this invention proposes a scattering power composite feature. vari-azi Its expression is

[0143] Feature vari-azi =(P R +P H )×P V (17)

[0144] Based on the above analysis of scattering characteristics, it can be concluded that Feature vari-azi The feature value is larger at targets with changing orientations, such as vehicles. However, for ground backgrounds such as roads, grass, and trees, the rotational dihedral scattering and spiral scattering are weak, thus their feature values ​​are smaller. vari-azi The value is relatively small. Furthermore, the introduction of volume scattering power further suppresses ground background elements such as roads and grasslands, which have dominant surface scattering mechanisms.

[0145] S213. Combining Feature para-azi and Feature vari-azi This invention proposes a feature for the recombination of scattering power. vehicle Its expression is

[0146]

[0147] S22. A composite feature detector for scattering power was constructed to achieve effective detection of PolSAR vehicle targets.

[0148] S22 includes the following specific steps:

[0149] S221. Feature Based on Scattering Power Recombination Characteristics vehicle Construct a scattering power composite feature detector Det vehicle As shown in the following formula:

[0150] Det vehicle =Density(Feature) vehicle (19)

[0151] Where Density(X) represents the density of X, which is the mean value of X in the N×N neighborhood of the center pixel, and the size of the neighborhood is set to 3×3;

[0152] S222. Set the detection threshold to TH vehicle The PolSAR vehicle target detection results are as follows:

[0153]

[0154] The method for determining the detection threshold in S222 includes the following steps:

[0155] S2221. Select a typical ground background region as training data and calculate Det within that region. vehicle The value;

[0156] S2222. To eliminate the influence of outliers on threshold selection, the obtained Det vehicleThe values ​​are sorted in ascending order, and the largest values, which account for 0.1% to 0.5% of the total number of pixels in the training data, are removed. The remaining pixels' Det values ​​are then... vehicle The maximum value is determined to be TH vehicle .

[0157] Experimental results

[0158] The experimental data used in this invention consists of two Ku-band fully polarized SAR images obtained from an airport by a UAV-borne SAR system: UAV-borne SAR data 1 and UAV-borne SAR data 2. These two images represent the same target scene at different headings and observation angles. The imaging scenes mainly include vehicle targets, grassland, bare ground, roads, and trees. The original azimuth and range resolutions are both 0.15 meters. To suppress the influence of speckle noise, the data underwent a 3×3 fine Lee filter. Figure 2 Pauli maps, optical maps, and ground truth maps of vehicle targets are provided for these two datasets. The ground truth maps of vehicle targets were manually labeled based on the optical and Pauli maps. UAV-borne SAR data 1 contains 7 vehicle targets (V1-V7, where V3-V5 are large buses and the rest are cars), and UAV-borne SAR data 2 contains 3 vehicle targets (V1-V3, where V1 is a medium-sized truck and V2 and V3 are cars). Figure 2 In the image, (a) is the Pauli map of UAV-borne SAR data 1; (b) is the optical map of UAV-borne SAR data 1; (c) is the ground truth map of vehicle targets in UAV-borne SAR data 1; (d) is the Pauli map of UAV-borne SAR data 2; (e) is the optical map of UAV-borne SAR data 2; and (f) is the ground truth map of vehicle targets in UAV-borne SAR data 2.

[0159] Fine five-component polarization decomposition was performed on UAV-borne SAR data 1 and UAV-borne SAR data 2. The RGB pseudo-color images and each scattering component of the decomposition results are shown below. Figure 3 and Figure 4 In the RGB pseudo-color image, the red channel represents scattering from man-made targets (even-order scattering, spiral scattering, and dihedral scattering), the green channel represents volume scattering (multiple interactive scattering), and the blue channel represents surface scattering. Figure 3 and Figure 4 It can be observed that vehicle targets with parallel azimuth exhibit strong surface scattering, even-order scattering, and multiple interactive scattering, and their RGB pseudo-color images show red or magenta (V1, V3-V5, and V7 in UAV-borne SAR data 1, and V1 and V2 in UAV-borne SAR data 2). In contrast, vehicle targets with variable azimuth exhibit more significant rotational dihedral scattering, helical scattering, and multiple interactive scattering, and their RGB pseudo-color images show orange-red or yellow (V2 and V6 in UAV-borne SAR data 1, and V3 in UAV-borne SAR data 2).

[0160] To quantitatively analyze the polarization decomposition results, this invention selected three ground background regions (ROI 1 to ROI 3, representing grassland, road, and trees, respectively) from UAV-borne SAR data 1 and UAV-borne SAR data 2, and statistically analyzed the average power proportion of each scattering component at the vehicle target and the ground background, as shown in Tables 1 and 2. Tables 1 and 2 show that for azimuth-parallel vehicle targets (V1, V3-V5, and V7 in UAV-borne SAR data 1, and V1 and V2 in UAV-borne SAR data 2), the average power proportions of even-order scattering and surface scattering components are relatively high, with their sum exceeding 80%. The average power proportion of the rotating dihedral scattering component is very small, not exceeding 2%. For variable-azimuth vehicle targets (V2 and V6 in UAV-borne SAR data 1, and V3 in UAV-borne SAR data 2), the average power proportion of the even-order scattering component is significantly lower than that of azimuth-parallel vehicle targets, while the average power proportions of multiple interactive scattering, spiral scattering, and rotating dihedral scattering components are significantly higher. For grassland (ROI 1 in UAV-borne SAR data 1 and UAV-borne SAR data 2), the average power of its surface scattering component accounts for the largest proportion, approximately 40%, indicating that its dominant scattering mechanism is surface scattering. Roads (ROI 2 in UAV-borne SAR data 1 and UAV-borne SAR data 2) exhibit very weak backscattering, appearing darker on the RGB pseudo-color map. Besides the surface scattering component of the road surface, the interaction between the road surface and roadside trees, guardrails, etc., generates numerous even-order scattering or multiple interactive scattering. For trees (ROI 3 in UAV-borne SAR data 1 and UAV-borne SAR data 2), in addition to the volume scattering component of the canopy, the even-order scattering component generated by the dihedral structure formed by the trunk and the ground, or the surface scattering component of the ground beneath the tree, are also relatively strong. Based on the above analysis, the scattering power composite feature in equation (18) can be inferred. vehicle It can improve the contrast between the vehicle target and the ground background, which is helpful for subsequent vehicle target detection.

[0161] Table 1. Average power percentage (%) of each scattering component in UAV-borne SAR data.

[0162]

[0163]

[0164] Table 2. Average power percentage of each scattering component in UAV-borne SAR data (%)

[0165]

[0166] Vehicle target detection performance and comparison: Based on the results of fine five-component polarization decomposition, this invention constructs a scattering power composite feature detector (Det). vehicle This method is used for vehicle target detection. Dai's method, NPNF method, and Cloude's method are selected as comparison methods, and the target-to-clutter ratio (TCR), receiver operating characteristic curve (ROC), and figure of merit (FoM) are selected as evaluation indicators. The expressions for TCR and FoM are as follows:

[0167] TCR = 10log 10 (S T / S C )

[0168]

[0169] Among them, S T and S C N represents the power of the vehicle target and the background ground clutter, respectively. tt N fa and N mt N represents the number of positive detections, the number of false alarms, and the number of missed detections, respectively. gt =N tt +N mt This indicates the actual number of vehicle targets.

[0170] Figure 5 and Figure 6 Amplitude plots for different detectors are presented on UAV-borne SAR data 1 and UAV-borne data 2, respectively. Figure 5 and Figure 6 The two sets of figures clearly show that at targets with parallel orientations, all detectors exhibit larger amplitude values. However, at targets with varying orientations, compared to other vehicle detectors, the Det detector proposed in this invention... vehicle It exhibits more significant amplitude values ​​while preserving the outline and detail information of the vehicle target. Furthermore, Figure 7 The TCR variation of different detectors at various vehicle targets is shown, where the power S of the vehicle target is... T The mean of the detector amplitude at the vehicle target pixel, and the background clutter power S C The mean value of the detector amplitude is selected for the clutter region. The distribution of vehicle targets in UAV-borne SAR data 1 and UAV-borne data 2 is shown in [reference needed]. Figure 2 (c) and Figure 2 The vehicle target ground truth map in (f) shows the selected clutter region as... Figure 2 (a) and Figure 2The areas marked with red boxes in (d) are all 100 pixels × 100 pixels in size. Figure 7 It can be seen that Det vehicle It can significantly improve the TCR at vehicle targets with varying azimuth. For V2 and V6 in UAV-borne SAR data 1, it can improve the TCR by more than 6 dB compared to other detectors. For V3 in UAV-borne SAR data 2, it improves the TCR by approximately 4 dB compared to Dai's method and the NPNF detector. And for parallel-azimuth vehicle targets, Det... vehicle The obtained TCR is generally close to or slightly higher than that of other detectors. Therefore, the Det proposed in this invention... vehicle It can enhance the contrast between vehicle targets, especially vehicles that change direction, and the ground background.

[0171] ROC curves are commonly used to evaluate the detection performance of different detectors, and therefore this invention also uses them as one of the evaluation metrics. Since the number of vehicle targets in UAV-borne SAR data 1 and UAV-borne SAR data 2 is relatively small, to make the ROC curves smoother, this invention treats each pixel of the vehicle target as an independent target. Therefore, the detection probability P... d The false alarm probability P is the proportion of correctly detected target pixels to the total number of target pixels. fa This represents the proportion of background pixels that are incorrectly detected as targets. For different detectors, the detection threshold is varied, and the corresponding detection probability P is calculated. d And the false alarm probability P fa Thus, its ROC curve is obtained as follows: Figure 8 As shown. From Figure 8 It can be seen that, under the same false alarm probability, the Det of this invention... vehicle It has a higher detection probability P d Its ROC curve is at the top. Therefore, by Figure 8 As can be seen from the ROC curve, Det vehicle Its detection performance is superior to other detectors.

[0172] Figure 9 and Figure 10 The detection results of different methods on UAV-borne SAR data 1 and UAV-borne SAR data 2 are presented respectively, where green, yellow, and red rectangles represent correctly detected targets, missed targets, and false alarm targets, respectively. For Det vehicle Its threshold TH vehicle The method for determining the clutter training area is described in step S222. Figure 2 The areas marked with red boxes in (a) and (d) are shown. Both the NPNF method and Cloude's method use CFAR detection: (1) Selecting the areas with Det vehicle(2) Set a reasonable false alarm probability P. fa The corresponding detection threshold and preliminary detection results are obtained. To reduce false alarms in the preliminary detection results, this invention uses area filtering for post-processing to obtain the final detection results. Dai's method, based on the MPWF CFAR detection results, combines Wishart classification and superpixel segmentation processing to obtain the final vehicle target detection results.

[0173] from Figure 9 As can be seen, for UAV-borne SAR data 1, the method of this invention can correctly detect all vehicle targets without generating false alarms. In contrast, the NPNF method and Cloude's method both missed detecting the changing-direction vehicle target V6, and all three comparative methods generated one false alarm. The location of the false alarm is compared with... Figure 2 The optical image in (b) reveals that the road sign standing to the left of the road at the bottom of the image is parallel to the radar's flight direction, and its backscattering generates strong energy, causing a false alarm in this area using the contrast method. For UAV-borne SAR data 2, from... Figure 10 It can be seen that all methods can detect all vehicle targets, but the detection results given by the method of this invention and Dai's method do not contain any false alarms, while the NPNF method and Cloude's method generated 2 and 3 false alarms, respectively. (Comparison) Figure 10 The test results of (d) and Figure 2 As shown in the optical image in (e), the metal guardrail along the roadside produces strong backscattering, leading to numerous false alarms in Cloude's method. While Dai's method achieves good detection results, it requires complex Wishart classification and superpixel segmentation processing on the MPWF CFAR results to remove artificial clutter interference. The method of this invention utilizes Det... vehicle By setting a detection threshold to obtain preliminary detection results and performing simple area filtering post-processing, excellent vehicle target detection results can be achieved.

[0174] The specific FoM (Fine Factors of Quality) for different methods are shown in Table 3.

[0175] Table 3. FoM values ​​for different detection methods

[0176]

[0177]

[0178] As can be seen from Table 3, the method of the present invention has the highest FoM on both UAV-borne SAR data 1 and UAV-borne data 2, therefore its detection performance is optimal. Furthermore, Figure 9 and Figure 10 Enlarged images of the detection results of different methods at vehicles in varying orientations are provided, specifically in the areas marked with magenta boxes. It can be seen that, compared to other methods, the method of this invention can retain more contour and detail information of vehicles in varying orientations, thanks to Det... vehicle It has a larger TCR at targets with changing orientation.

[0179] Therefore, compared to the comparative methods, the PolSAR vehicle target detection method based on fine polarization decomposition of this invention can significantly improve the TCR at vehicle targets with varying orientations, exhibiting the optimal ROC curve and the highest FoM, demonstrating excellent detection performance. Furthermore, the method of this invention can retain more contour and detail information of the vehicle target, which is helpful for further applications such as target classification and recognition.

[0180] Finally, it should be noted that although the present invention has been described with reference to specific embodiments, those skilled in the art should recognize that the above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Various equivalent changes or substitutions can be made without departing from the concept of the present invention. Therefore, any changes or modifications to the above embodiments within the essential spirit of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A PolSAR vehicle target detection method based on fine polarization decomposition, characterized in that, Includes the following steps: S1. Construct a rotating dihedral scattering model and propose a fine five-component polarization decomposition method to perform fine polarization decomposition on the scattering of PolSAR vehicle targets; S2. Based on robust scattering power recombination characteristics, a scattering power recombination characteristic detector is constructed to achieve effective detection of PolSAR vehicle targets; S2 includes the following steps: S21. A simple and physically clear composite feature of scattering power was constructed. S21 includes the following steps: S211. For vehicle targets with parallel azimuth, i.e., targets pointing parallel to or perpendicular to the radar's flight direction, a composite scattering power characteristic is proposed. Its expression is as follows: ； S212. For vehicles with variable azimuth, i.e., targets whose direction of travel is neither parallel nor perpendicular to the radar's flight direction, a composite scattering power characteristic is proposed. Its expression is as follows: ； Used to construct composite features and The power of each scattering component has been normalized, and the expression is as follows: ， in, , These correspond to surface scattering power, even-order scattering power, volume scattering power, spiral scattering power, and rotational dihedral scattering power, respectively. The range of values ​​is ; S213. Combination and The composite characteristics of scattering power were proposed. Its expression is as follows: ； S22. A composite feature detector for scattering power was constructed to achieve effective detection of PolSAR vehicle targets.

2. The PolSAR vehicle target detection method based on fine polarization decomposition according to claim 1, characterized in that, S1 includes the following steps: S11. For vehicle target detection tasks, a rotating dihedral scattering model RDSM was constructed based on the generalized rotating dihedral scattering model GRDSM. S12. A refined five-component polarization decomposition method is proposed by combining the rotating dihedral angle scattering model RDSM.

3. The PolSAR vehicle target detection method based on fine polarization decomposition according to claim 2, characterized in that, S11 includes the following steps: S111. A reasonable modeling of scattering from a rotating dihedral structure is proposed. Based on the cross-scattering model (CSM), a generalized rotating dihedral scattering model (GRDSM) is derived, with its coherence matrix... The expression is as follows: in, , And there are ; S112. Based on the generalized rotating dihedral scattering model GRDSM, a rotating dihedral scattering model RDSM is constructed, whose coherence matrix is... The expression is as follows: ， In the above formula The expression is derived from the eigenvalues ​​of the coherence matrix and is as follows: in, Represents the total scattered power. i , i=1,2,3 represent the eigenvalues ​​of the coherence matrix.

4. The PolSAR vehicle target detection method based on fine polarization decomposition according to claim 3, characterized in that, The coherence matrix of the cross-scattering model CSM in S111 The expression is as follows: in, This indicates the dominant polarization azimuth angle.

5. The PolSAR vehicle target detection method based on fine polarization decomposition according to claim 2, characterized in that, S12 includes the following steps: S121. The expression for the refined five-component polarization decomposition method is as follows: in, Represents the measurement coherence matrix. , , , and These correspond to surface scattering, even-order scattering, volume scattering, spiral scattering, and rotating dihedral scattering models, respectively. , , , and These represent the corresponding expansion coefficients. , , and The expression is as follows: ； S122. Measure the coherence matrix Expanding by elements, we obtain the following system of equations: S123. Using branch conditions With some model parameters of the equation system fixed, when At that time, it was assumed that the dominant scattering mechanism of the remaining matrix was surface scattering, and let ,and Otherwise, even-order scattering is considered the dominant scattering mechanism of the remaining matrix, and let ,and ; S124. In non-rotational dihedral regions or other natural regions, the equations in the system of equations The values ​​are ignored during the calculation, and the same applies to the dihedral region of rotation. The solution for all model parameters is as follows: Therefore, the power of each scattering component is as follows: 。 6. The PolSAR vehicle target detection method based on fine polarization decomposition according to claim 1, characterized in that, S22 includes the following steps: S221. Based on the recombination characteristics of scattering power Construct a scattering power composite feature detector As shown in the following formula: in, express The density, that is At the center pixel The mean of the neighborhood, the size of the neighborhood is set to ; S222. Set the detection threshold to... The PolSAR vehicle target detection results are as follows: 。 7. The PolSAR vehicle target detection method based on fine polarization decomposition according to claim 6, characterized in that, The detection threshold in S222 The method for determining it includes the following steps: S2221. Select a typical ground background area as training data, and calculate the calculation results within this area. The value; S2222. The obtained The values ​​are sorted in ascending order, and 0.5% of the total number of pixels in the training data are removed. The remaining pixels are... The maximum value is determined as .

Images