A method for attitude estimation of non-cooperative targets in space

Abstract

The present invention discloses a method for estimating the attitude of a space non-cooperative target, comprising the following steps: S1: acquiring an image of a space non-cooperative target; S2: identifying target angular information from the target image based on error ellipse and Hough transform; S3: using Hu's invariant moment to describe the target attitude image, and then using a pattern classifier to identify the pitch angle; S4: based on the target angular information and target machine axis information obtained in step S2 and step S3, using a projective set method to reconstruct the three-dimensional attitude of the target. Aiming at the characteristics of rich close-range imaging information in the terminal guidance stage of infrared seekers, the present invention conducts research on guidance information mining technology, and uses image pattern recognition and computer vision methods to identify the new guidance information of aircraft target attitude, which is of great significance to improving the intelligent level and combat effectiveness of my country's infrared guidance technology and achieving leapfrog development.

Description

Technical Field

[0001] The present invention relates to the technical field of navigation, and in particular to a method for estimating the posture of a spatial non-cooperative target. Background Art

[0002] Infrared-guided air-to-air missiles, a key branch of air-to-air missiles, are the primary weapon for close-range combat. Currently, the fourth-generation missiles currently under development or already in combat in countries around the world all utilize precision guidance technologies such as infrared imaging guidance. Infrared imaging guidance, along with thrust vectoring and high-speed, low-drag aerodynamic configurations, have become the defining technologies of fourth-generation air-to-air combat missiles. Infrared imaging guidance combines infrared imagers with image processing, using digital signal processing to analyze images. This enables missiles to directly impact targets or hit critical (vulnerable) areas of a target, making it a highly efficient and cost-effective guidance technology.

[0003] During a missile's flight, target-related information is acquired by the seeker. The relationship between the missile's control commands and the seeker information is generated by the guidance laws implemented within the missile system. Different guidance laws have varying dependencies on target parameters and missile flight parameters, and their adaptability to different attack conditions also varies significantly. With the advancement of modern air combat technology and the maturation of highly maneuverable advanced fighter aircraft, the requirements for missile guidance system accuracy are becoming increasingly stringent. Currently, the seekers of fourth-generation missiles worldwide utilize advanced image processing techniques to accurately track, identify, and engage targets. While tracking the target is a prerequisite for accurate interception, simply tracking the target is not enough. Modern air combat requires accurate targeting of vital areas of the target while maintaining lock-on status to maximize the probability of a kill.

[0004] The vast majority of currently published literature focuses on the detection and extraction of position-related parameters. This research is limited to target detection and interference analysis using infrared radiation and motion characteristics, resulting in a single attribute value for a point of interest. Furthermore, while extensive research has been conducted on aircraft and interference (jamming bombs, clouds, etc.), aircraft type identification, and advanced guidance law design, significant gaps remain in the field of three-dimensional attitude recognition of aircraft targets and the design of novel guidance laws that utilize attitude angle information. Summary of the Invention

[0005] In response to the above problems, the present invention aims to provide a method for estimating the attitude of a non-cooperative target in space, which uses image pattern recognition and computer vision methods to identify the attitude of the target aircraft. This method is of great significance to improving the intelligent level and combat effectiveness of my country's infrared guidance technology and achieving leapfrog development.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A method for estimating the posture of a non-cooperative target in space, characterized by comprising the following steps:

[0008] S1: Acquire images of non-cooperative targets in space;

[0009] S2: Based on the error ellipse and Hough transform, the target direction angle information is identified from the target image;

[0010] S3: Use Hu's invariant moment to describe the target posture image, and then use the pattern classifier to identify the pitch angle;

[0011] S4: Based on the target direction angle information and target machine axis information obtained in steps S2 and S3, the three-dimensional posture of the target is reconstructed using the projective set method.

[0012] Furthermore, the specific operation of step S2 includes the following steps:

[0013] S21: Filter and segment the target image to obtain a binary image;

[0014] S22: Use morphological opening operation to remove small lines and burrs in binary images;

[0015] S23: Based on the error ellipse theory, calculate the target image point distribution and obtain the major and minor axes and corresponding angles of the error ellipse;

[0016] S24: Design a cropping template image, based on the angles of the two main axes of the error ellipse obtained in step S23 and The template image is rotated and cropped, and an AND operation is performed on it with the original target image to obtain two thin strip area images consisting only of pixels near the long and short principal axes, which are also cropped images.

[0017] S25: extracting the main axes of the two cropped images using the Hough transform method, thereby obtaining the two longest main axes;

[0018] S26: Use logical reasoning to confirm the main axis where the target aircraft head is located, and output the direction angle φ corresponding to the target aircraft head.

[0019] Furthermore, the specific operation of step S26 includes the following steps:

[0020] S261: Based on the two main axes obtained by the Hough transform method, determine the distances between the four endpoints of the two main axes and the centroid, denoted as {L1, L2, S1, S2}, where {L1, L2} represent the distances between the two endpoints corresponding to the major axis L and the centroid, L1 + L2 = L; {S1, S2} represent the distances between the two endpoints corresponding to the minor axis S and the centroid, S1 + S2 = S;

[0021] S262: If max{L1, L2, S1, S2} = max{L1, L2} does not hold, the nose is considered to be on the minor axis, and max{S1, S2} corresponds to the position of the nose;

[0022] S263: If max{L1, L2, S1, S2} = max{L1, L2}, and at the same time The nose is considered to be on the minor axis, and max{S1, S2} corresponds to the position of the nose; where th is the length allowable threshold;

[0023] S263: If max{L1, L2, S1, S2}=max{L1, L2} holds, but If it does not hold, the nose is considered to be on the long axis, and max{L1, L2} corresponds to the position of the nose.

[0024] Furthermore, the target aircraft axis information in step S3 includes a yaw angle φ and a pitch angle θ; the method for identifying the yaw angle φ and the pitch angle θ specifically includes the following steps:

[0025] S31: Identify the yaw angle using a method for identifying the direction angle in the nose plane or other characteristic quantities that can reflect the yaw angle;

[0026] S32: Use Hu's invariant moment to describe the target posture image, and then use the pattern classifier to identify the pitch angle.

[0027] Furthermore, the specific operation of step S3 includes the following steps:

[0028] S31: converting the real-time target posture image obtained by infrared imaging into a binary target image through image segmentation;

[0029] S32: extracting the target region center moment vector from the binarized target image, and obtaining Hu's seven invariant moments or calculating the 12-dimensional RTS invariant of the target region center extension;

[0030] S33: normalizing the Hu moment vector calculated in step S32;

[0031] S34: Construct a training dataset using the normalized Hu moments obtained under different posture conditions, build a three- or four-layer neural network structure, train the neural network classifier using the Levenberg-Marquardt learning algorithm, and initialize the network weights and thresholds using the Nguyen-Widrow algorithm;

[0032] S35: The normalized Hu moment vector obtained in step S33 is input into a neural network classifier, and the angle between the main axis of the aircraft space and the imaging plane, i.e., the pitch angle θ, is obtained by fitting using the BP or RBF neural network approximation capability.

[0033] Furthermore, the specific operation of step S4 includes the following steps:

[0034] S41: Establish the missile body coordinate system MXYZ and the target projection plane coordinate system oxyz. Mo is the missile-target line. CD is used to represent the spatial axis. EF is the line connecting the two ends of the corresponding target lateral wing. The projection of the axis in the xoy plane is IJ. The target coordinate system is mx T y T z T , z T Assuming that the fuselage normal is upward, the wing tip connecting line and y T The axes coincide with each other, and the intersection with the XMZ plane is H;

[0035] S42: Let q φ and q θ The angle between the normal of the target aircraft and the plane CDM is defined as γ, and the angle between the normal of the target aircraft and the plane ABO is defined as γ. TM , from which the azimuth angle φ of the target body relative to the missile body coordinate system can be determined TM 、Elevation angle θ TM and roll angle γ TM ;

[0036] S43: The three Euler angles of the target aircraft relative to the image coordinate system are obtained through the imaging system, and the attitude of the target aircraft relative to the missile can be obtained;

[0037] S44: Add the target aircraft's attitude relative to the missile to the missile-target distance R to determine the target aircraft's spatial attitude relative to the missile;

[0038] S45: Based on the rotation matrix of the missile body coordinate system relative to the ground coordinate system, and the spatial position of the target aircraft relative to the missile, the yaw, pitch and roll angles of the target aircraft relative to the ground coordinate system can be derived.

[0039] The beneficial effects of the present invention are:

[0040] Aiming at the characteristics of rich close-range imaging information in the terminal guidance stage of infrared seekers, the present invention conducts research on guidance information mining technology, and adopts image pattern recognition and computer vision methods to identify the new guidance information of aircraft target posture, filling the gap in my country's research on three-dimensional posture recognition of aircraft targets and design of new guidance laws using posture angle information. It is of great significance to improving the intelligence level and combat effectiveness of my country's infrared guidance technology and achieving leapfrog development. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of identifying target direction angle information from a target image based on error ellipse and Hough transform in the present invention;

[0042] Figure 2 Schematic diagram of the error ellipse structure in the present invention;

[0043] Figure 3 This is a schematic diagram of the main axis detection of the simple error ellipse in the present invention;

[0044] Figure 4 Schematic diagram of the cropping template and cropping image in the present invention;

[0045] Figure 5 It is the logical reasoning diagram of the present invention;

[0046] Figure 6 This is a flowchart of the target machine axis attitude estimation process in the present invention;

[0047] Figure 7 This is the invariant moment translation rotation scaling invariance experiment of the present invention;

[0048] Figure 8 Reconstructing the three-dimensional attitude coordinate system of the aircraft using the projective geometry method of the present invention;

[0049] Figure 9 This is the experimental result of (30°)-(30°)-(30°) in the simulation experiment 1 of the present invention;

[0050] Figure 10 The experimental results of (-90°)-(75°)-(15°) in the simulation experiment 1 of the present invention are shown below:

[0051] Figure 11 This is a three-dimensional diagram of the recognition error distribution in case (1) of the simulation experiment of the present invention;

[0052] Figure 12 This is a two-dimensional half-graph of the recognition error distribution in case (1) of the simulation experiment of the present invention;

[0053] Figure 13 This is a three-dimensional diagram of the recognition error distribution in case (2) of the simulation experiment 1 of the present invention;

[0054] Figure 14 This is a two-dimensional half-graph of the recognition error distribution in case (2) of the simulation experiment 1 of the present invention;

[0055] Figure 15 The training error distribution results when the network weights and thresholds are initialized using the Nguyen-Widrow algorithm and the network is trained using the Levenberg-Marquardt learning algorithm in case (1) of the second simulation experiment of the present invention;

[0056] Figure 16 In case (1) of the second simulation experiment of the present invention, the network weights and thresholds are initialized using the Nguyen-Widrow algorithm, and the error distribution results are tested when the network is trained using the Levenberg-Marquardt learning algorithm;

[0057] Figure 17 For the case (1) of the second simulation experiment of the present invention, a generalized regression neural network is used to estimate the posture angle and the network training error distribution result is trained by the Levenberg-Marquardt learning algorithm;

[0058] Figure 18 In the case (1) of the second simulation experiment of the present invention, a generalized regression neural network is used to estimate the posture angle and the Levenberg-Marquardt learning algorithm is used to train the network to test the error distribution result;

[0059] Figure 19 The training error distribution result of the training sample feature set constructed using the extended twelve-dimensional invariant moment in case (2) of the second simulation experiment of the present invention;

[0060] Figure 20 The error distribution results of the training sample feature set test are constructed using the extended twelve-dimensional invariant moment in case (2) of the second simulation experiment of the present invention;

[0061] Figure 21 In case (2) of the second simulation experiment of the present invention, a generalized regression neural network is used to estimate the distribution result of the attitude angle training error;

[0062] Figure 22 In case (2) of the second simulation experiment of the present invention, a generalized regression neural network is used to estimate the distribution results of the attitude angle test error. DETAILED DESCRIPTION

[0063] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0064] A method for estimating the posture of a spatial non-cooperative target comprises the following steps:

[0065] S1: Acquire an image of a non-cooperative target in space. Specifically, acquire an infrared image of the target through the seeker.

[0066] S2: Based on the error ellipse and Hough transform, the target direction angle information is identified from the target image;

[0067] S3: Use Hu's invariant moment to describe the target posture image, and then use the pattern classifier to identify the pitch angle;

[0068] S4: Based on the target direction angle information and target machine axis information obtained in steps S2 and S3, the three-dimensional posture of the target is reconstructed using the projective set method.

[0069] Specifically, the specific operation process of step S2 is as shown in the attached Figure 1 As shown, the following steps are included:

[0070] S21: Filter and segment the target image to obtain a binary image;

[0071] S22: Use morphological opening operation to remove small lines and burrs in binary images;

[0072] S23: Based on the error ellipse theory, calculate the target image point distribution and obtain the major and minor axes and corresponding angles of the error ellipse;

[0073] Specifically, in two-dimensional space, a set of observed words usually always falls within an elliptical area around its expected value with a probability of 1-α (α is the confidence). The diffusion ellipse can also be called the error ellipse, as shown in the attached figure. Figure 2 As shown in , if the expected value of the observation is regarded as the diffusion center of the ellipse, and the radius vector from the center to any point on the circumference is regarded as the potential difference in that direction, then the major and minor semi-axes of the ellipse represent the extreme values ​​of the potential difference.

[0074] Treating all points on the image as observations, the maximum and minimum values ​​of the error ellipse constructed from these observations correspond to the major and minor axes of the target. This is the theoretical basis for using error ellipse theory to determine the orientation angle of an image target.

[0075] After the original image is preprocessed by image filtering, segmentation, etc., the coordinates p of the target area point (or target edge point) are extracted. i (x i ,y i )(i=1,2,…n), where n is the number of points, the coordinates of the image centroid are

[0076]

[0077] Center the lattice coordinates and construct the elements in the real symmetric matrix D

[0078]

[0079] Formula (2) can be used to calculate the major and minor axis direction angles of the point distribution. Wei Yi et al. used the error ellipse theory to identify aircraft and non-aircraft targets. Their idea was to construct a matrix judgment domain in an extreme state based on the two main factors that affect the aircraft shape and its digital characteristics: the thickness of the fuselage and the angle between the wing and the fuselage. According to mathematical statistics, the area within this rectangular domain is the confidence region to be judged. Experiments have proved the feasibility of the scheme.

[0080] S24: Design and crop template image;

[0081] Specifically, the major and minor axes obtained by simply using the error ellipse are sometimes not the expected major and minor axes, such as Figure 3 The images shown are for the (30°)-(30°)-(30°) attitude and the (-90°)-(75°)-(15°) attitude. The red line represents the major axis, and the blue line represents the minor axis. The (30°)-(30°)-(30°) attitude captures the correct major axis, but the major axis is not the principal axis of the nose. In the (-90°)-(75°)-(15°) attitude image, due to the large out-of-plane rotation angle, the projection of the aircraft's longitudinal axis is identified as the minor axis, and the axis of the two wings is considered the major axis. This can lead to two problems: 1) The major and minor axes obtained using the error ellipse method cannot determine which axis is the nose axis, and the four axis endpoints cannot determine which axis represents the nose position; 2) Due to changes in the aircraft's attitude, the major and minor axes obtained from the error ellipse are likely not the axis of the nose.

[0082] The cropping template image is the same size as the original image and is formed by performing a pixel-by-pixel AND operation on two semi-circular images. The only parameter of the cropping template is the vertex angle θ, which is determined by the error detected by the error ellipse principal axis detection method when the posture changes. It is generally set to 30° to 40°.

[0083] According to the angles of the two main axes of the error ellipse obtained in step S23 and The template image is rotated and cropped, and an AND operation is performed with the original target image to obtain two thin strip area images consisting only of pixels near the long and short main axes, namely the cropped images, as shown in the attached figure. Figure 4 shown.

[0084] S25: extracting the main axes of the two cropped images using the Hough transform method, thereby obtaining the two longest main axes;

[0085] Specifically, the two cropped images obtained in step S24 contain the main axes of the aircraft. The main axes of the two cropped images are extracted using the Hough transform method introduced above, thereby obtaining the two longest main axes.

[0086] S26: Use logical reasoning to confirm the main axis where the target machine head is located, and output the direction angle φ corresponding to the target machine head. The logical reasoning block diagram is shown in the attached figure. Figure 5 As shown, the specific operations include:

[0087] S261: Based on the two main axes obtained by the Hough transform method, determine the distances between the four endpoints of the two main axes and the centroid, denoted as {L1, L2, S1, S2}, where {L1, L2} represent the distances between the two endpoints corresponding to the major axis L and the centroid, L1 + L2 = L; {S1, S2} represent the distances between the two endpoints corresponding to the minor axis S and the centroid, S1 + S2 = S;

[0088] S262: If max{L1, L2, S1, S2} = max{L1, L2} does not hold, the nose is considered to be on the minor axis, and max{S1, S2} corresponds to the position of the nose;

[0089] S263: If max{L1, L2, S1, S2} = max{L1, L2}, and at the same time The nose is considered to be on the minor axis, and max{S1, S2} corresponds to the position of the nose; where th is the length allowable threshold, which is set to 0.1-0.2 based on experience.

[0090] S263: If max{L1, L2, S1, S2}=max{L1, L2} holds, but If it does not hold, the nose is considered to be on the long axis, and max{L1, L2} corresponds to the position of the nose.

[0091] Further, the specific operation process of step S3 is as shown in the attached Figure 6 As shown, the following steps are included:

[0092] S31: converting the real-time target posture image obtained by infrared imaging into a binary target image through image segmentation;

[0093] Real-time infrared images obtained by infrared imaging are converted into binary target images through image segmentation. Due to varying environmental conditions, the grayscale within the target often varies significantly, which is not a stable image feature. Therefore, the algorithm design ignores this variation. However, the aircraft tail flame contains attitude information, which can be processed separately in practical applications.

[0094] S32: extracting the target region center moment vector from the binarized target image, and obtaining Hu's seven invariant moments or calculating the 12-dimensional RTS invariant of the target region center extension;

[0095] Practice has shown that directly using the origin moment or central moment as an image feature cannot guarantee that the feature is simultaneously invariant to translation, rotation, and scale. In fact, if only the central moment is used to represent the image feature, the feature is only invariant to translation; if the normalized central moment is used, the feature is not only invariant to translation but also invariant to scale, but not to rotation. It seems that directly using the origin moment or central moment is not enough to achieve simultaneous invariance to translation, rotation, and scale transformations. To this end, M. Hu first proposed invariant moments in 1961. He gave the definition of continuous function moments and the basic properties of moments, proved the properties of moments such as translation invariance, rotation invariance, and scale invariance, and specifically gave expressions for seven invariant moments that are invariant to translation, rotation, and scale.

[0096] The seven invariant moments are composed of linear combinations of the second-order and third-order central moments, and the specific expressions are as follows:

[0097] M1=η 20 +η 02 (3)

[0098]

[0099] M3=(η 30 -3η 12 ) 2 +(3η 21 -η 03 ) 2 (5)

[0100] M4=(η 30 +3η 12 ) 2 +(3η 21 +η 03 ) 2 (6)

[0101] M5=(η 30 -3η 12 )(η 30 +η 12 )[(η 30 +η 12 ) 2 -3(η 12 +η 03 ) 2 ]+(3η 21 -η 03 )(η 21 -η 03 )[3(η 30 +η 12 ) 2 -(η 21 +η 03 )2 ] (7)

[0102] M6=(η 20 -or 02 )[(or 30 +n 12 ) 2 -(or 21 -or 03 ) 2 ]+4th 11 (or 30 +n 12 )(or 21 -or 03 ) (8)

[0103] M7=(3rd 12 -or 30 )(or 30 +n 12 )[(or 30 +n 12 ) 2 -3(h 21 +n 03 ) 2 ]+(3rd 21 -or 03 )(or 21 +n 03 )[3(h 03 +n 12 ) 2 -(or 12 +n 03 ) 2 ] (9)

[0104] Standardized center distance pq defined as

[0105]

[0106] Among them,

[0107] m pqis the (p+q)th order central moment of f(x, y). Among the seven moment invariants, the first six are mirror invariant, while M7 is -M7 under mirror transformation. They are suitable for describing the overall shape of the target, and therefore have a wide range of applications in edge extraction, image matching, and target recognition. The computational complexity of these seven Hu moments is different, and the information content is also different. The useful information in the image is generally concentrated in the low-order moments with relatively small computational complexity, while the high-order moments have a large computational complexity and contain some detailed information that is easily interfered by noise. The differences between the high-order moments of each target are not easy to distinguish. In the experiment, all 7 Hu invariant moment features were used as the feature vectors of the image. Forming a feature space (M1, M2, M3, M4, M5, M6, M7).

[0108] Furthermore, in some applications, it is often difficult to distinguish multiple complex image patterns based solely on the existing seven Hu invariant moments (such as requiring the classification and recognition of handwritten samples of more than 6,000 commonly used Chinese characters). This requires us to seek more transformation invariants that reflect image information based on the construction rules of invariant moments and do not rely on the seven classical moments.

[0109] Liu Jin, Zhang Tianxu and others equated the proof of moment invariance to the proof of angular parameter invariance of the corresponding rotationally invariant trigonometric functions, proposed the concept of invariant moment space, summarized the general construction law of invariant moments, and derived five new moment expressions (M8, M9, M 10 , M 11 , M 12 ), other higher-order invariant moment expressions can be constructed using similar schemes. Furthermore, a general formula for higher-order invariant moment expressions is derived using the trigonometric production method. Experiments demonstrate that these new invariant moment expressions exhibit excellent invariance properties.

[0110] The expanded five invariant moments are

[0111] M8=2{η 11 [(η 30 +η 12 ) 2 -(η 21 +η 03 ) 2 ]-(η 20 -η 02 )(η 30 +η 12 )(η 21 +η 03 )} (11)

[0112] M9=[(η 30 -3η 12 )(η 30 +η 12 )+(3η 21 -η03 )(η 21 +η 03 )](η 20 -η 02 )+2η 11 [(3η 21 -η 03 )(η 30 +η 12 )-(η 30 -3η 12 )(η 21 +η 03 )] (12)

[0113] M 10 =[(3η 21 -η 03 )(η 30 +η 12 )-(η 30 -3η 12 )(η 21 +η 03 )](η 20 -η 02 )-2η 11 [(η 30 -3η 12 )(η 30 +η 12 )+] (13)

[0114] M 11 =(η 04 +η 40 -6η 22 ) 2 +16(η 31 -η 13 ) 2 (14)

[0115]

[0116] Among them, M8, M9, M 10 They are all moment expressions with a total degree of 3, M 11 , M 12 is a fourth-order invariant moment, and M7, M8 and M 10 Both have the ability to distinguish mirror images.

[0117] These invariants are proven in the continuous case. For discrete images in practical applications, scaling and rotation may produce some distortion. The experimentally derived invariants are not strictly invariant to geometric transformations of the original image, but rather fluctuate within a certain range. Numerous experiments have shown that the instability of discrete image invariants increases with the degree of the invariant polynomial and the order of the moment.

[0118] Figure 7 Shown are four images that have undergone translation, rotation, and scaling transformations. (a) is the original image; (b) is the original image magnified 1.5 times and rotated 45°; (c) is the original image reduced 0.8 times and rotated 120°; and (d) is the original image magnified 1.3 times and rotated -90°. While there are no legends for translation transformations, their invariance is readily apparent. Table 1 shows that all moment values ​​remain essentially constant after translation, scaling, and rotation of the original image. Due to discretization, the individual feature quantities fluctuate within a certain range, but this fluctuation is not significant. Experiments demonstrate the effectiveness of these 12-dimensional invariant moments. Increasing the dimension improves the shape representation of the target pattern and increases pattern recognition accuracy.

[0119] Table 1 Experimental data of RTS invariance of invariant moments

[0120] Moment invariants Figure 7 (a) Figure 7 (b) Figure 7 (c) Figure 7 (d) <![CDATA[M1]]> 1.43 1.4366 1.427 1.4323 <![CDATA[M2]]> 4.5502 4.5544 4.5424 4.5648 <![CDATA[M3]]> 5.9183 5.931 5.8838 5.9564 <![CDATA[M4]]> 7.8397 7.8871 7.7844 7.8596 <![CDATA[M5]]> 14.85 14.941 14.733 14.879 <![CDATA[M6]]> 10.825 10.902 10.743 10.795 <![CDATA[M7]]> 15.453 15.487 15.412 15.572 <![CDATA[M8]]> 10.253 10.294 10.201 10.3 <![CDATA[M9]]> 9.3175 9.3439 9.2804 9.3559 <![CDATA[M 10 ]]> 10.285 10.33 10.21 10.314 <![CDATA[M 11 ]]> 6.0381 6.0928 6.0125 6.0429 <![CDATA[M 12 ]]> 8.7131 8.6891 8.7238 8.7367

[0121] S33: normalizing the Hu moment vector calculated in step S32;

[0122] Data normalization is the process of scaling data so that it falls within a small, specific range. Since the units of measurement for each indicator in the credit index system are different, in order to include the indicators in the evaluation calculation, they need to be normalized and their values ​​mapped to a certain range through functional transformation. Before training and testing the BP network, the Hu moment vector needs to be normalized to prevent neuron output saturation caused by excessively large net input absolute values.

[0123] S34: Construct a training dataset from the normalized Hu moments obtained under different posture conditions, build a three- or four-layer neural network structure, train the neural network classifier using the Levenberg-Marquardt learning algorithm, and initialize the network weights and thresholds using the Nguyen-Widrow algorithm;

[0124] S35: The normalized Hu moment vector obtained in step S33 is input into a neural network classifier, and the angle between the main axis of the aircraft space and the imaging plane, i.e., the pitch angle θ, is obtained by fitting using the BP or RBF neural network approximation capability.

[0125] Since there is ambiguity in the attitude angle recognition results of aircraft imaging, the hypothesis testing method can be used to determine the aircraft attitude angle of the current frame image.

[0126] Furthermore, the specific operation of step S4 includes the following steps:

[0127] S41: Establish the missile body coordinate system MXYZ and the target projection surface coordinate system oxyz, as shown in the attached Figure 8As shown, Mo is the missile-target line, CD is used to represent the spatial machine axis, EF is the line connecting the two ends of the corresponding target lateral wing, the projection of the machine axis in the xoy plane is IJ, and the target machine coordinate system is Tx T y T z T , z T Assuming that the fuselage normal is upward, the wing tip connecting line and y T The axes coincide with each other, and the intersection with the XMZ plane is H;

[0128] S42: Let q φ and q θ The angle between the normal of the target aircraft and the plane CDM is defined as γ, and the angle between the normal of the target aircraft and the plane ABO is defined as γ. TM , from which the azimuth angle φ of the target body relative to the missile body coordinate system can be determined TM 、Elevation angle θ TM and roll angle γ TM ;

[0129] S43: The three Euler angles (azimuth angle φ, pitch angle θ, and roll angle γ) of the target aircraft relative to the image coordinate system are obtained through the imaging system to obtain the attitude of the target aircraft relative to the missile;

[0130] S44: Add the target aircraft's attitude relative to the missile to the missile-target distance R to determine the target aircraft's spatial attitude relative to the missile;

[0131] S45: Based on the rotation matrix of the missile body coordinate system relative to the ground coordinate system, and the spatial position of the target aircraft relative to the missile, the yaw, pitch and roll angles of the target aircraft relative to the ground coordinate system can be derived.

[0132] Simulation experiment 1:

[0133] A simulation experiment was conducted on a method based on error ellipse and Hough transform to identify target direction angle information from target image. The results are shown in the attached figure. Figure 9 and attached Figure 10 As shown in the figure, each set of images represents the original image, two cropped images superimposed with Hough line detection, and the original image with two detected principal axes superimposed. As can be seen from the figure, our algorithm effectively avoids two problems: the error ellipse algorithm mistakenly identifies the minor axis as the principal axis of the machine head, and the principal axis obtained by the statistical lattice is not the true principal axis.

[0134] Furthermore, this method was applied to full-pose images for experimental verification. First, the recognition probabilities for two special cases, one with the roll angle fixed at 0° and the other with the pitch angle fixed at 0°, were examined. Then, the recognition probability for all poses was examined. The results are as follows.

[0135] (1) Only the azimuth and pitch angles change, and the roll angle is fixed at 0°

[0136] The simulation sample data was obtained by selecting a sampling interval of every 15°. The range of the azimuth angle was -165° to 180°, and the range of the pitch angle was -75° to 75°. At this time, there were 24×11=264 sets of image data. When the sampling interval was 5°, the range of the azimuth angle was -175° to 180°, and the range of the pitch angle was -85° to 85°. At this time, there were 72×35=2520 sets of data. The corresponding experimental error statistics are shown in Table 2 and Appendix. Figure 11 、 12 Compared with the Hough transform method, the accuracy is significantly improved.

[0137] Table 2 Recognition probability under different error levels in case 1

[0138] Error level |Error|<10° |Error|<5° |Error|<3° Recognition probability (interval 15°) 95.45％ 87.50％ 72.73％ Recognition probability (5° interval) 90.2％ 81.59％ 68.53％

[0139] (2) Only the azimuth and roll angles change, and the pitch angle is fixed at 0°

[0140] When the pitch angle is fixed at 0° and the azimuth and roll angles are varied at intervals of 15°, 24×24=576 sample images can be simulated. When the azimuth and roll angles are sampled at intervals of 5°, 72×72=5184 image samples can be obtained. The corresponding experimental error statistics are shown in Table 3 and the attached table. Figure 13 、 14 shown.

[0141] Table 3 Recognition probability under different error levels in case 2

[0142] Error level |Error|<10° |Error|<5° |Error|<3° Recognition probability (interval 15°) 95.31％ 92.88％ 77.43％ Recognition probability (5° interval) 92.59％ 90.30％ 71.89％

[0143] (3) Recognition probability under all posture conditions

[0144] The algorithm's experimental results were examined when the azimuth, pitch, and roll angles were all changing. A total of 24 × 11 × 24 = 6336 images were collected, and the experimental results are shown in Table 4. The average processing time per image was 567.3 ms. Three sets of experiments showed that this method not only significantly improved the recognition probability at various error levels, correctly distinguishing the long and short principal axes of the nose and wings, but also accurately determined the nose position, representing a significant improvement over the method of simply extracting the principal axes using the Hough transform.

[0145] Table 4 Recognition probability under different error levels in case three

[0146] Error level |Error|<10° |Error|<5° |Error|<3° Recognition probability (interval 15°) 73.03％ 61.05％ 50.06％

[0147] Simulation experiment 2:

[0148] In the second simulation experiment, in order to verify the effectiveness of the present invention's use of Hu's invariant moment to describe the target posture image and then use the pattern classifier to perform pitch angle recognition algorithm, the F16 aircraft posture database image was tested and simulated, mainly using Hu moment and extended invariant moment to construct feature vectors, and using BP and RBF neural network classifiers to realize the estimation of pitch angle θ. Due to the ambiguity of posture in aircraft imaging, images with different posture angles may be the same. If the training data obtained from imaging at all angles is brought into the neural network training, the network training will not converge due to sample incompatibility, and the generalization ability will be very poor. Therefore, complete non-conflicting samples should be selected, and the image posture sequence hypothesis testing method can solve this problem. Therefore, in this simulation experiment, only the posture angle investigation range is selected.

[0149] The training database is built using a scene-driven method based on real-world shooting and 3D models, with pitch and roll angles spaced 5° apart.

[0150] The resulting image dataset contains 19 × 37 = 703 binary images of the F-16 aircraft. The test database is generated at 7° intervals, totaling 13 × 26 = 338 images. The experimental platform is Matlab 6.5, a Core 2 Duo CPU, a T8100, and 2GB of RAM. All images have a resolution of 200 × 200 pixels.

[0151] (1) Hu moment construction of feature vector recognition

[0152] The training sample feature set was constructed using Hu's seven invariant moments, and a simple BP network structure with 7 inputs, 10 hidden layers, and 1 output layer was selected to estimate the main shaft pitch angle. The network weights and thresholds were initialized using the Nguyen-Widrow algorithm, and the network was trained using the Levenberg-Marquardt learning algorithm. The training time was 68.7660s on the Matlab 6.5 platform, and the training was repeated 4000 times. The error was 0.00445006 when the training stopped, and the test time was 0.015s. The average recognition error within 20° was 3.472°. The distribution of the training and test errors is shown in the attached figure. Figure 15 and attached Figure 16 The error distribution probability and statistical characteristics are shown in Tables 5 and 6. Then, a generalized regression neural network is used to estimate the posture angle. The network is trained using the Levenberg-Marquardt learning algorithm with a network structure of 7-703-1. The training time is 0.1880s and the test time is 0.015s on the Matlab 6.5 platform. The mean error within the test error of 20° is 2.7372°. The training and test error distributions are shown in Tables 5 and 6. Figure 17 and Figure 18 The error distribution probability and statistical characteristics are shown in Tables 7 and 8.

[0153] Table 5. Training and testing error probabilities

[0154] Recognition rate |Error|<10° |Error|<5° |Error|<3° |Error|>20° Training accuracy 99.573％ 87.909％ 68.279％ 0 Test accuracy 95.858％ 76.627％ 57.101％ 0.29586％

[0155] Table 6. Training and test error statistics

[0156] Absolute error distribution mean median variance Training accuracy 2.5071° 1.9455° 4.4409 Test accuracy 3.5584° 2.6094° 12.4648

[0157] Table 7. Training and testing error probabilities

[0158] Recognition rate |Error|<10° |Error|<5° |Error|<3° |Error|>20° Training accuracy 99.858％ 96.586％ 92.319％ 0 Test accuracy 94.083％ 83.136％ 69.231％ 0.88757％

[0159] Table 8. Training and test error statistics

[0160] Absolute error distribution mean median variance Training accuracy 0.7727° 0.0058583° 2.2272 Test accuracy 2.9889° 1.914° 15.1187

[0161] (2) Extending the invariant moment to construct the eigenvector

[0162] The extended 12-dimensional invariant moment was used to construct a training sample feature set. A simple network structure with 12 inputs, 15 hidden layers, and 1 output layer was selected to estimate the main axis pitch angle. The training time was 261.2643 seconds, and the training was repeated 4000 times. The error at the end of the training was 0.00005647, and the test time was 0.016 seconds. The distribution of the training and test errors is shown in the figure below. Figure 19 、 Figure 20 , as shown in Table 9 and Table 10.

[0163] Table 9. Training and test error statistics

[0164] Recognition rate |Error|<10° |Error|<5° |Error|<3° |Error|>20° Training accuracy 100％ 100％ 98.578％ 0 Test accuracy 95.858％ 91.42％ 81.657％ 2.367％

[0165] Table 10. Training and test error statistics

[0166] Absolute error distribution mean median variance Training accuracy 0.71623° 0.52191° 0.48257 Test accuracy 3.8245° 1.1338° 279.9904

[0167] The following generalized regression neural network is used to estimate the posture angle. The network structure is 12-703-1. The training time is 0.1560s and the test time is 0.141s under the Matlab6.5 platform. The training and test error distributions are as follows: Figure 21 、 Figure 22 , as shown in Table 11 and Table 12.

[0168] Table 11. Training and test error statistics

[0169] Recognition rate |Error|<10° |Error|<5° |Error|<3° |Error|>20° Training accuracy 99.858％ 98.009％ 94.452％ 0 Test accuracy 98.817％ 92.604％ 84.024％ 0

[0170] Table 12. Training and test error statistics

[0171] Absolute error distribution mean median variance Training accuracy 0.4937° 0.00022495° 1.544 Test accuracy 1.8874° 1.1447° 4.5765

[0172] The four sets of experiments above demonstrate that the neural network fitting method achieves high accuracy and real-time performance in estimating the pitch angle between the aircraft axis and the imaging plane. Comparing the two networks, although the BP network achieved the same probability of accuracy within 10° as the generalized regression neural network when using the 7-dimensional invariant moment, the probability of accuracy within 5° and within 3° was lower than that of the GRNN network. Therefore, the GRNN network has better function approximation capabilities. Furthermore, the GRNN network is very fast to train, with training time almost comparable to the test time. When using the same network, the 12-dimensional invariant moment generally outperformed the 7-dimensional invariant moment, indicating that increasing the dimension of the feature vector appears to enhance the descriptive power of features and achieve better recognition results. The GRNN network and the 12-dimensional invariant moment achieved the highest test accuracy, achieving an error within 10° with a probability of 98.817% and an error within 3° with a probability of 84.024%. Both the median and mean errors were within 2°, demonstrating excellent performance.

[0173] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for estimating a spatial non-cooperative target posture, characterized in that: The following steps are included: S1: Acquire images of non-cooperative targets in space; S2: Based on the error ellipse and Hough transform, the target direction angle information is identified from the target image; S3: Use Hu's invariant moment to describe the target posture image, and then use the pattern classifier to identify the pitch angle; S4: Based on the target direction angle information and pitch angle information obtained in step S2 and step S3, the three-dimensional posture of the target is reconstructed using the projection set method; The specific operation of step S2 includes the following steps: S21: Filter and segment the target image to obtain a binary image; S22: Use morphological opening operation to remove small fine lines and burrs in binary images; S23: Based on the error ellipse theory, the target image dot distribution is calculated to obtain the major and minor axes and corresponding angles of the error ellipse; S24: Design a cropping template image, based on the angles of the two main axes of the error ellipse obtained in step S23 and The template image is rotated and cropped, and an AND operation is performed with the original target image to obtain two thin strip area images consisting only of pixels near the long and short main axes, namely the cropped images; S25: extracting the main axes of the two cropped images using the Hough transform method, thereby obtaining the two longest main axes; S26: using logical reasoning to confirm the main axis where the target aircraft head is located, and outputting the direction angle φ corresponding to the target aircraft head; The specific operation of step S3 includes the following steps: S31: converting the real-time target posture image obtained by infrared imaging into a binary target image through image segmentation; S32: extracting the target region center moment vector from the binarized target image, obtaining the seven invariant moments of Hu or calculating the 12-dimensional RTS invariant of the target region center extension; S33: normalizing the Hu moment vector calculated in step S32; S34: construct a training data set from the normalized Hu moments obtained under different posture conditions, build a three- or four-layer neural network structure, train the neural network classifier using the Levenberg-marquardt learning algorithm, and use the Nguyen-Widrow algorithm to initialize the network weights and thresholds; S35: The normalized Hu moment vector obtained in step S33 is input into a neural network classifier, and the angle between the main axis of the aircraft space and the imaging plane, that is, the pitch angle θ, is obtained by fitting using the BP or RBF neural network approximation capability.

2. A method for estimating a spatial non-cooperative target posture according to claim 1, characterized in that: The specific operation of step S26 includes the following steps: S261: According to the two main axes obtained by the Hough transform method, the distances between the four endpoints of the two main axes and the centroid are determined, which are recorded as {L1, L2, S1, S2}, where {L1, L2} respectively represent the distances between the two endpoints corresponding to the long axis L and the centroid, L1+L2=L; {S1, S2} respectively represent the distances between the two endpoints corresponding to the short axis S and the centroid, S1+S2=S; S262: If max{L1,L2,S1,S2}=max{L1,L2} does not hold, it is considered that the nose is on the minor axis, and max{S1,S2} corresponds to the position of the nose; S263: If max{L1, L2, S1, S2} = max{L1, L2} holds, and at the same time It is considered that the nose is on the minor axis, and max{S1,S2} corresponds to the position of the nose; where th is the length allowable threshold; S263: If max{L1,L2,S1,S2}=max{L1,L2}, but If it is not true, the nose is considered to be on the long axis, and max{L1,L2} corresponds to the position of the nose.

3. A method for estimating a spatial non-cooperative target posture according to claim 2, characterized in that: The specific operation of step S4 includes the following steps: S41: Establish the missile body coordinate system MXYZ and the target projection plane coordinate system oxyz. Mo is the missile-target line. CD is used to represent the spatial machine axis. EF is the line connecting the two ends of the corresponding target lateral wing. The projection of the machine axis in the xoy plane is IJ. The target machine coordinate system is Tx T y T z T , z T Assume that the fuselage is normal to the top and the wing tip line is T The axes coincide with each other, and the intersection with the XMZ plane is H; S42: Let q φ and q θ The angle between the normal direction of the target aircraft and the plane CDM is defined as γ, and the angle between the normal direction of the target aircraft and the plane ABO is defined as γ. TM , from which the azimuth angle φ of the target body relative to the missile body coordinate system can be determined TM 、Height angle θ TM and the roll angle γ TM ; S43: The three Euler angles of the target aircraft relative to the image coordinate system are obtained through the imaging system, so as to obtain the attitude of the target aircraft relative to the missile; S44: Add the target aircraft's attitude relative to the missile to the missile-target distance R to determine the target aircraft's spatial attitude relative to the missile; S45: Based on the rotation matrix of the missile body coordinate system relative to the ground coordinate system, and the spatial position of the target aircraft relative to the missile, the yaw, pitch and roll angles of the target aircraft relative to the ground coordinate system can be derived.

Images