A multi-view radar image ATR optimal observation path planning method

Abstract

The application discloses a multi-view radar image ATR optimal observation path planning method, which is applied to the technical field of radar imaging and aims at the problem that the radar imaging platform of the prior art is difficult to explore in an unknown environment. According to actual task requirements, a geometric model and an optimization mathematical model are built. Based on a basic architecture of a deep convolutional neural network, a multi-view network architecture integrated by multiple base classifiers is constructed in combination with the characteristics of multi-view input, so that automatic target recognition of multi-view radar images is realized. Based on an artificial potential field path planning algorithm, the best viewpoint selection of the radar image ATR is realized in combination with the characteristics of the flight of the radar imaging platform in the unknown environment. The best path planning of the flight of the radar imaging platform is realized, and the constraint condition of the optimal recognition performance is also met.

Description

Technical Field

[0001] This invention belongs to the field of radar imaging technology, and specifically relates to a radar automatic target recognition and path planning technology. Background Technology

[0002] Automatic target recognition (ATR) of radar images is of great significance for mineral resource exploration, marine monitoring, and military-related fields. While single-input radar image ATR algorithms have been extensively studied, their performance varies depending on the viewpoint and the type of radar image. Multi-view radar images contain richer target information than single-view images, thus multi-view radar image ATR can achieve better recognition performance.

[0003] Multi-view radar image ATR selects the optimal viewpoint or perspective for identification and plans an appropriate detection path to achieve the best recognition performance. Some studies have shown that the achievements in radar image ATR path planning are effective. Most perspective views are implemented under known global environmental information, such as the multi-view radar image ATR based on network ensemble and graph search proposed in the literature "Jifang Pei, Yulin Huang, WeiboHuo, Yuan Xue, Yin Zhang, and Jianyu Yang, 'Multi-view radar ATR based on networks ensemble and graph search, in 2018 IEEE Radar Conference (RadarConf18), 2018, pp.0355–0360,' which can find the optimal multi-view radar image ATR flight path under known conditions. However, in actual detection, it is difficult to obtain environmental information in advance.

[0004] To accomplish the target recognition task, it is necessary to find the optimal viewpoint or perspective in an unknown environment and independently plan the best observation path. To achieve autonomous path planning for multi-view radar image ATR in unknown environments, corresponding geometric and mathematical models were established. To solve for the optimal solution of the objective function of the mathematical model, the Artificial Potential Field (APF) method was used to solve for the optimal flight trajectory. Therefore, the path search algorithm solves the problem of selecting the optimal viewpoint and observation path in multi-view radar image ATR. The multi-view radar images are acquired onto the solved optimal viewpoint, and the final recognition result is obtained through ensemble analysis using base classifiers. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes an optimal observation path planning method for multi-view radar image ATR. By constructing an optimization mathematical model based on task requirements and performing optimization solutions, the optimal observation path planning for the radar imaging platform is achieved, while simultaneously realizing the best recognition performance.

[0006] The technical solution adopted in this invention is: a method for optimal observation path planning of multi-view radar image ATR, comprising:

[0007] S1. The multi-view radar target observation path planning problem is modeled as a constrained multi-objective optimization problem;

[0008] S2. Construct multiple two-view convolutional neural network base classifiers to decompose the radar image sequence into a series of radar image pairs; train the constructed multiple two-view convolutional neural network base classifiers based on the radar image pairs.

[0009] S3. The output results of the trained two-view convolutional neural network base classifiers are fused to obtain the target recognition result of the multi-view radar image.

[0010] S4. Place the radar imaging platform in an abstract artificial gravitational field, and plan the next imaging viewpoint area starting from the initial viewpoint.

[0011] S5. Calculate the attractive and repulsive forces from each viewpoint in the next imaging viewpoint region to the current viewpoint;

[0012] S6. Consider angle constraints;

[0013] S7. Based on steps S5 and S6, determine the location where the current radar imaging platform experiences the maximum net force as the next imaging position of the radar imaging platform.

[0014] S8. Obtain the optimal imaging viewpoint and form the observation path by interpolation between imaging viewpoints.

[0015] The beneficial effects of this invention are as follows: Based on actual task requirements, this invention constructs a geometric model and an optimal mathematical model; based on the basic architecture of a deep convolutional neural network and combined with the characteristics of multi-view input, it constructs a multi-view network architecture that integrates learning from multiple base classifiers to achieve automatic target recognition of multi-view radar images; based on an artificial potential field path planning algorithm and combined with the characteristics of radar imaging platforms flying in unknown environments, it achieves the optimal viewpoint selection for radar image ATR; this invention not only achieves optimal path planning for radar image ATR in unknown environments, but also achieves the best recognition performance. Attached Figure Description

[0016] Figure 1 A flowchart of the method of the present invention.

[0017] Figure 2 Geometric model diagram used in specific embodiments of the present invention;

[0018] Among them, (a) is the three-dimensional geometric model of the multi-view radar image ATR; (b) is the top view of the multi-view radar image ATR geometric model.

[0019] Figure 3 The specific embodiments of this invention employ a deep learning framework for multi-view radar image target classification and recognition.

[0020] (a) is the ensemble learning framework; (b) is the basic classifier network structure.

[0021] Figure 4 The stress diagram of the radar imaging platform used in the specific embodiment of the present invention at the current location.

[0022] Figure 5 A schematic diagram of corner constraints used in a specific embodiment of the present invention.

[0023] Figure 6 The three-view, four-view, and five-view path planning diagrams obtained through specific embodiments of the present invention;

[0024] Among them, (a) is a three-view flight path map; (b) is a four-view flight path map; (c) is a five-view flight path map; (d) is the first four-view flight path exploration process; (e) is the second four-view flight path exploration process; and (f) is the third four-view flight path exploration process.

[0025] Figure 7 Bar charts showing the recognition rates of three-view, four-view, and five-view perspectives obtained through specific embodiments of the present invention;

[0026] Among them, (a) is the three-view recognition rate; (b) is the four-view recognition rate; and (c) is the five-view recognition rate. Detailed Implementation

[0027] This invention is primarily verified using simulation experiments. All steps and conclusions were verified correctly using PyCharm on the Windows 10 operating system platform. To facilitate understanding of the technical content of this invention by those skilled in the art, the following description, in conjunction with accompanying drawings, further elaborates on the invention.

[0028] S1. Constructing the geometric model and the optimal mathematical model. Based on the current detection environment and mission requirements, given the starting point and number of imaging angles of the radar imaging platform, and considering a series of constraints such as terrain, threats, mission requirements, and aircraft maneuverability, a multi-view radar image target autonomous detection and recognition model is constructed. A deep learning multi-view radar image target recognition framework is established, and an autonomous observation path planning scheme for unknown environments is proposed. The optimal imaging angle position is selected, and interpolation is performed between adjacent imaging angle positions to form the flight path of the radar image platform, achieving an effective balance between recognition performance and unknown environment detection during multi-view radar image ATR. The flowchart of this invention is shown below. Figure 1 As shown.

[0029] In multi-view radar image target recognition tasks, the platform detects and images the target from different perspectives along a planned path, acquiring a series of multi-view radar images. These multi-view radar images are then input into the ATR system to obtain the target's category attributes. The actual multi-view radar target recognition task mode is shown in the appendix. Figure 2 As shown.

[0030] The selection of radar platform imaging location also needs to consider the optimization and balance of many factors such as threat area and platform flight path distance cost. Therefore, the multi-view radar target observation path planning problem is modeled as a multi-objective optimization problem with constraints. The optimal mathematical model formula is as shown in formula (1):

[0031]

[0032] in, For the viewing angle range φ s and Recognition rate;

[0033]

[0034]

[0035]

[0036]

[0037] t ii This indicates the threat value between imaging viewpoint locations;

[0038]

[0039]

[0040] The objective function P(·) characterizes the recognition performance of the multi-view radar image target recognition system; a larger function value indicates a higher recognition rate. The objective function T(·) characterizes the safety of the radar platform during the detection and recognition process; a smaller T(·) value indicates a lower risk to the system. The objective function R(·) characterizes the flight path cost of the radar platform during the detection and recognition process; a smaller R(·) value indicates a lower flight cost. In planning the observation path, relevant constraints should also be considered, such as the platform's flight dynamics characteristics, flight path boundaries, and imaging point location limitations. In this optimization model, the radar platform's flight dynamics characteristic θ(U) is limited by the parameter λ. The set Θ represents the search space for solving the imaging point location, further standardizing the radar platform's imaging perspective. In this embodiment, λ is set to 110.

[0041] S2. Constructing a multi-view ensemble learning convolutional neural network architecture. Based on the ensemble learning convolutional neural network scheme and combined with the requirements of target recognition and classification tasks, a deep learning framework for multi-view radar image target classification and recognition is established. A reasonable multi-view classifier is designed to spontaneously learn the optimal classification information from a limited number of radar image target samples from different angles, compensating for the fluctuations and omissions of information in single-view imaging, and achieving efficient and accurate target recognition. The multi-view radar image target classification and recognition deep learning framework established in this invention is as follows: Figure 3 As shown.

[0042] First, based on the imaging of the radar imaging platform at different angles, the radar image sequence is decomposed into a series of radar image pairs according to the arrangement and combination of different angles. Then, three two-view convolutional neural network base classifiers are designed. The radar image pairs are input into specific base classifiers. The number of training and validation samples for the three base classifiers is shown in Table 1 in the appendix. The number of base classifiers is also determined according to the arrangement and combination of different angles.

[0043] Table 1. Number of training and validation samples for the three basic classifiers.

[0044]

[0045] The raw SAR image obtained by the SAR sensor is {x1,x2,…,x n}, corresponding imaging angle Based on the flight feasible region division criteria in the target recognition geometric model, the sector region can be represented as a set of view intervals. Therefore, any original SAR image x i′ It belongs to only one viewpoint range. These raw SAR images can be recombined into a dataset. in N sThis represents the number of combinations after recombination of samples. Any base classifier CNN#l only works on the set of view intervals. The SAR image pairs are classified. Here, l represents the viewpoint interval difference, 1 ≤ i ≤ N. s , if i+l≤N s If δ = 0, then δ = 0; otherwise, δ = -1. Therefore, the SAR dataset X can be reassembled into a set of SAR image pairs. A subset of data is randomly selected from these SAR image pairs as training samples for the base classifier CNN#1. The remaining SAR image pairs serve as the validation set, used to evaluate the recognition performance of the base classifier CNN#1 across different viewpoint ranges.

[0046] The corresponding SAR image training sample set generated by the above method is used to train a series of basic classifiers. Based on the ensemble learning theory, the final recognition result of the ensemble multi-view classifier is shown in Equation (2).

[0047]

[0048] Where h represents the final recognition result of the integrated multi-view classifier, {x1,x2,…,x k} represents the input multi-view radar image sequence, p i (y|x i ,x i+1 ) is the base classifier CNN#l for conditions The C-dimensional posterior probability vector is obtained by classifying radar image pairs. Where 1 ≤ s ≤ N. s Let l = 1, 2, ..., L, where L is the number of base classifiers. ∏(·) represents the index of the element containing the maximum value in the input vector.

[0049] The posterior probability vector output by the base classifier can be represented as:

[0050]

[0051] In formula (3), the elements of the vector represent the posterior probabilities of the categories identified by the base classifier for the input SAR image, where

[0052] p i (y j |x i ,x i+1 ) = p i (y j ∣z) (4)

[0053]

[0054] Where z is the input of the Softmax classifier in the i-th convolutional neural network.

[0055] S3. Update the detection viewpoint set. Based on the basic idea of ​​the artificial potential field method, the radar imaging platform is placed in an abstract artificial gravitational field. Starting from the initial viewpoint, the next imaging viewpoint region is planned. The set of the next imaging viewpoint regions is shown in formula (6). Let the current location of the radar imaging platform be U. i Point, the set of the next possible imaging locations is

[0056]

[0057] in Let L represent the possible next imaging location, and r represent the number of basic classifiers and r be the number of flyable rings of the radar imaging platform. The force diagram of the radar imaging platform at its current position is shown below. Figure 4 As shown.

[0058] S4. Update the gravitational, repulsive, and resultant forces at the current detection viewpoint. Calculate the gravitational and repulsive forces from each viewpoint in the next imaging viewpoint region to the current viewpoint, and define the gravitational force formula as shown in formula (7):

[0059]

[0060] Where, k att As a correction factor, ρ(U) i U j () represents the imaging position U i with U j The recognition rate between them is obtained from the previous two-view convolutional neural network classifier. The repulsion formula is as shown in formula (8):

[0061]

[0062] Where r(U) i U j () represents the imaging position U i with U j The Euclidean distance between them, t(U) i U j () represents the imaging position U j For U i The threat value. Further calculation of the resultant force between the two is given by formula (9):

[0063] F sum (U i U j ) = F att (U i U j )+F rep (U i U j(9)

[0064] S5. Update Constraints. In many practical situations, due to the dynamic performance of the radar imaging platform, it is necessary to consider angle constraints. The current viewpoint and the previous viewpoint are assumed to be U. i =(x i ,y i H) and U i-1 =(x i-1 ,y i-1 ,H), where (x i ,y i ) and (x i-1 ,y i-1 () represents the coordinates, and H represents the flight altitude of the radar imaging platform. Then, according to U i+1 To U i The resultant force sequence traversal U i+1 The set of detection viewpoints. Calculate the angle between the three viewpoints. The angle formula is as shown in formula (10):

[0065]

[0066] The detection viewpoint set here is the set of the next imaging viewpoint regions in step S3.

[0067] A schematic diagram of the current position and angle constraints of the radar imaging platform is shown below. Figure 5 As shown.

[0068] S6. Determine the next detection viewpoint position. Based on steps S4 and S5, determine the point where the resultant force on the current radar imaging platform is the maximum as the next imaging position of the radar imaging platform. Repeat steps S3, S4, S5, and S6 until all viewpoints have been selected.

[0069] S7. Optimal Path Planning. Based on the above methods, the optimal imaging viewpoint is obtained. The observation path is formed by interpolation between imaging viewpoints, ultimately realizing path planning for radar image ATR in unknown environments.

[0070] The flight path results for three-view, four-view, and five-view perspectives are shown below. Figure 6 As shown, from Figure 6 As can be seen, the method of this invention can plan the optimal imaging point and the optimal observation path for target recognition tasks with different starting points and different numbers of viewing angles, effectively achieving an optimized balance between recognition performance and detection efficiency. Using this invention, a high target recognition rate for multi-view radar images can be obtained. As the number of input viewing angles increases, the target recognition rate for multi-view radar images will also be significantly improved. The recognition rate histogram is shown below. Figure 7As shown in Table 2 of the appendix, the average recognition rate is as follows. Experimental results show that by selecting the optimal imaging point using the artificial potential field method and integrating it with a convolutional neural network, the optimal classification information of radar image targets can be discovered from different perspectives, achieving efficient and accurate identification of multi-view radar image targets in threatening unknown environments.

[0071] Table 2 Average Recognition Rate from Multiple Viewpoints

[0072]

[0073] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A method for optimal observation path planning in multi-view radar image ATR, characterized in that, include: S1. The multi-view radar target observation path planning problem is modeled as a constrained multi-objective optimization problem; S2. Construct multiple two-view convolutional neural network base classifiers to decompose the radar image sequence into a series of radar image pairs; train the constructed multiple two-view convolutional neural network base classifiers based on the radar image pairs. S3. The output results of the trained two-view convolutional neural network base classifiers are fused to obtain the target recognition result of the multi-view radar image. S4. Place the radar imaging platform in an abstract artificial gravitational field, and plan the next imaging viewpoint area starting from the initial viewpoint. S5. Calculate the attractive and repulsive forces from each viewpoint in the next imaging viewpoint region to the current viewpoint; The gravity mentioned in step S5 is obtained based on the target recognition results of the multi-view radar image in step S3; the gravity formula is: ； in, For correction factor, For imaging position and The recognition rate between them; The repulsive force described in step S5 is calculated based on the Euclidean distance between the imaging viewpoints and the threat value; the repulsive force formula is: ； in, For imaging position and The Euclidean distance between them For imaging position right Threat value; S6. Consider angle constraints; the angle constraints mentioned in step S6 are specifically: ； in, Indicates the current imaging viewpoint position. Indicates the position of the previous imaging viewpoint. Indicates the position of the next imaging viewpoint. express arrive The vector, express arrive The vector, Represents the absolute value symbol. For parameters; S7. Based on steps S5 and S6, determine the location where the current radar imaging platform experiences the maximum net force as the next imaging position of the radar imaging platform. S8. Obtain the optimal imaging viewpoint and form the observation path by interpolation between imaging viewpoints.

2. The optimal observation path planning method for multi-view radar image ATR according to claim 1, characterized in that, The optimization problem described in step S1 is specifically as follows: ； objective function Characterizing the recognition performance of a multi-view radar image target recognition system, the objective function Characterizing the security of the radar platform during the detection and identification process, the objective function Characterizes the flight path cost of the radar platform during the detection and identification process; The radar platform's flight dynamics characteristics are given, where λ is a parameter constraint. The search space for solving the location of the imaging point. This represents the position of the imaging viewpoint.

3. The optimal observation path planning method for multi-view radar image ATR according to claim 2, characterized in that, Step S2 is based on the imaging of the radar imaging platform at different angles. According to the arrangement and combination of different angles, several view intervals are obtained. The radar image sequence is decomposed into a series of radar image pairs. The number of view intervals is equal to the number of basic classifiers of the two-view convolutional neural network. One basic classifier of the two-view convolutional neural network classifies only the radar image pairs of one view interval.

Images