Clustered fixed-wing aircraft trajectory estimation method based on timing similarity feature information

Abstract

The application discloses a cluster fixed-wing aircraft trajectory estimation method based on time sequence similar feature information, relates to the technical field of unmanned aerial vehicle cluster perception and deep learning, and first, a FractalAttNet fractal network model is designed for shallow feature extraction to enhance the shallow features and enrich the feature expression; secondly, a Few-Shot method is used, density estimation and dense sample estimation are performed on the original image by using independent samples inferred by the FractalAttNet fractal network, and the screening and resolution of the occluded and overlapped targets are realized; thirdly, according to the time sequence similar features, a multi-target matching method with high real-time performance is designed based on an auction algorithm, and the real-time performance requirement of multi-target matching is enhanced; finally, the observed radar data and trajectory information are re-encoded by using a Transformer method, and the trajectory prediction task of the flight target is completed by using time sequence correlation information.

Description

Technical Field

[0001] This invention relates to the fields of UAV swarm perception and deep learning technology, and in particular to a method for estimating the trajectory of swarm fixed-wing aircraft based on temporal similarity feature information. Background Technology

[0002] As a novel type of swarm system, drone swarms have a wide range of applications in both civilian and military fields. In the civilian sector, drone swarms have already seen initial applications in emergency rescue, precision agriculture, remote sensing and Earth observation, and light shows, with significant potential for future development. In military applications, drone swarms can perform a variety of combat missions, including reconnaissance and search, electronic warfare, area control, and saturation attacks.

[0003] With the advancement and iteration of swarm intelligence algorithms and multi-drone cooperative control algorithms, efficient data networks and information exchange have become the main constraints on the development of UAV swarms. Currently, communication methods based on satellite positioning and radio frequency networks suffer from susceptibility to interference, weak environmental adaptability, and data offset. Therefore, visual or passive radar has become a growing focus of attention.

[0004] Utilizing visual methods to achieve drone swarming offers significant advantages. First, visual information acquisition methods are most similar to biological swarming approaches, and their highly independent distributed architecture and scalability are essential for future large-scale drone swarming. Second, visual information acquisition methods can mitigate the lateral drift problem caused by satellite positioning, resulting in a higher degree of fit between the input and the decision-making module.

[0005] Using visual methods to identify and estimate the trajectory of swarm drones in large-scale scenarios is of great research significance. It can provide simple and effective data input for the decision-making and control modules of swarm drones, and lay the foundation for the realization of large-scale drone swarms.

[0006] Therefore, proposing a trajectory estimation method for clustered fixed-wing aircraft based on temporal similarity feature information to address the difficulties in existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides a trajectory estimation method for clustered fixed-wing aircraft based on temporal similarity feature information, which can perform well in most environments, is not affected by environmental and climatic conditions, improves the identification of dense small targets, and has high confidence and reliability.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] The trajectory estimation method for clustered fixed-wing aircraft based on temporal similarity feature information includes the following steps:

[0010] S1: Acquire first-person view image data of the cluster of aircraft;

[0011] S2: The small target recognition method based on shallow feature attention performs target recognition on the image data collected in S1 to obtain small target features;

[0012] S3: The dense target post-processing method based on feature matching performs density estimation and dense sample estimation on the small target features obtained in S2 to obtain the filtered small target features;

[0013] S4: The multi-target matching method based on the auction algorithm matches and distinguishes the small target features filtered in S3, thereby achieving matching and distinction of different small target features;

[0014] S5: Using an attention-based target trajectory estimation method, complete the estimation of the feature trajectories of small targets after matching and differentiation in S4.

[0015] The above method, optionally, uses the FractalAttNet fractal network model in the small target recognition method based on shallow feature attention in S2, which specifically includes two branches: feature extraction branch and filtering branch.

[0016] The above method, optionally, includes the following two branch details:

[0017] The feature extraction branch uses a bottleneck structure to complete feature extraction and utilizes the bottleneck structure to increase the depth of the shallow network.

[0018] The filtering branch uses the HourglassBlock module, which introduces the encoder-decoder structure of the target segmentation and key point recognition network to obtain small-range feature changes. The extracted shallow features are then input into the FractalAttNet fractal network model to complete the extraction of small target features after recognition.

[0019] The above method, optionally, can be expressed as the FractalAttNet fractal network model formula:

[0020]

[0021] Where SE represents the SELayer module, φ(x) represents the convolutional network, Att represents the mini-Hourglass detection module, and wl and w2 represent the feature weights learned by the DropPath method during training of the two branches, respectively.

[0022] Optionally, S3 of the above method includes: using the dense prediction module, after completing target recognition, taking the optimal output extracted from each layer as a reference sample, and using the extracted features as similar features, and using the feature matching method to directly complete the dense target segmentation and position estimation in the bottom layer domain.

[0023] Optionally, in the above method, S4 specifically involves matching the inter-frame feature relationships of the time sequence.

[0024] First, the recognition results between frames are compared, with the recognition result of the previous frame recorded in M ​​and the recognition result of the current frame recorded in N; then, the cost function is used for similarity matching to obtain the final cost matrix.

[0025] The cost function is:

[0026] C ij =||p i -p j ||2*(δS*w s )*lg((0.3δS+1)+1),δS=|S i -S j |·····(2)

[0027] Where p is the predicted center point, S is the predicted bounding box area, δS represents the change in bounding box volume between frames, i represents the frame number of the previous frame, j represents the frame number of the current frame, and w s The area represents the weight in the function, M represents the recognition result set of the previous frame, and N represents the recognition result set of the current frame.

[0028] To ensure the consistency of temporal features, the target extraction results are transformed and projected into an independent matrix position encoding.

[0029] The above method, optionally, includes the following specific content for S5:

[0030] Using the spatial correspondence of analytical geometry, the distance between machines and the size of the target bounding box exhibit a clear inverse relationship, and the calculation formula is as follows:

[0031]

[0032] Among them, s p f is the target pixel size. p s is the pixel focal length at the center point of the target. w d represents the actual target size, and d represents the relative distance.

[0033] Optionally, S5 may also include the following:

[0034] A preliminary distance estimation method for monocular cameras is obtained by directly training and converging the relevant parameters using the nonlinear least squares method.

[0035] Establish a negative exponential sample function model:

[0036]

[0037] in, Let x be the negative exponential sample function corresponding to the aircraft distance and bounding box, where a and b are optimization parameters and x is the current bounding box size.

[0038] The objective function obtained from the collected data target value is:

[0039]

[0040] Among them, y i This represents the actual distance value collected from the data, where m is the total number of data points collected.

[0041] Since the collected data are independent and identically distributed, and considering the nonlinear characteristics of the function, the Gauss-Newton method is used to perform a linear transformation on the objective function. The transformation relationship is as follows:

[0042]

[0043]

[0044]

[0045] Where J(x) is the Jacobian matrix of the residual r(x), Let T be the Heather matrix, and let T denote the transpose of the matrix. The first derivative of the function, The second derivative of the objective function;

[0046] Solve the linear equations using Newton's method using formulas (6), (7), and (8):

[0047]

[0048] The corresponding linear relationship is obtained:

[0049] J(x) T J(x)p GN =-J(x) T r(x)·····(10)

[0050] Where, p GN The linear value of the Gauss-Newton method is represented by the linear least squares method, which is used to optimize the objective function.

[0051] By using the Gauss-Newton method to solve for the corresponding values ​​of a and b in the objective function φ(x; a, b), the nonlinear constraint equations are obtained:

[0052]

[0053] A preliminary target distance estimation formula is obtained, and the projection rule is used to complete the preliminary estimation of the target position. Combined with S2, S3, and S4, the target bounding box is accurately extracted to complete the target distance estimation, which is then converted into a single-frame position estimation result.

[0054] Optionally, the above method may perform filtering and position optimization on the position estimation results; and convert the target position information into a visual radar matrix. The numerical coordinates in the matrix represent the polar coordinate projection position of the current target, and the matrix value represents the relative distance of the current target. Using the Transformer structure, a number is added to the value of each aircraft to distinguish different aircraft. The information of each frame is transformed into coded coordinates located in the same space R. The frame number information of consecutive frames is used as the position code input to the Transformer attention network for training to complete the trajectory estimation task of the target.

[0055] As can be seen from the above technical solution, compared with the prior art, the present invention provides a method for estimating the trajectory of a cluster of fixed-wing aircraft based on temporal similarity feature information:

[0056] (1) To improve the ability to distinguish dense targets with high repetition in clusters by using similar feature matching mechanism and to correct the non-maximum suppression part to suppress the false screening of overlapping targets, the recognition effect of dense small targets for cluster fixed-wing aircraft is effectively improved.

[0057] (2) Align the aircraft pose data and time series data, and use the newly proposed visual radar coding method and Transformer structure to complete the aircraft trajectory estimation and optimization tasks. Compared with data estimation methods such as Kalman filtering or polynomial fitting, it is more intuitive and has lower edge variability.

[0058] (3) By using visual methods, it is not subject to electromagnetic interference and can still work normally under strong electromagnetic interference. It can perform well in most environments and is not affected by environmental, climatic and other conditions.

[0059] (4) Using a deep learning-based target recognition algorithm has the advantages of high robustness, strong anti-interference and the ability to recognize multiple targets at the same time. In addition, the deep learning-based target recognition algorithm can meet the needs of embedded platforms after GPU acceleration. By enhancing shallow features and correcting the NMS method, the problem of recognizing dense small targets can be significantly improved.

[0060] (5) The visual radar coding method based on Transformer and the trajectory optimization method based on time information can effectively solve the jitter and offset problems of the current visual estimation results under the premise of low computing power, and have high confidence and reliability. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0062] Figure 1 A schematic diagram of an end-to-end network structure provided for existing technologies;

[0063] Figure 2 The collected data and annotation results provided by this invention are as follows: 2a represents the first-view image data of the swarm aircraft, and 2b represents the annotation results of 2a.

[0064] Figure 3 The network structure of the recognition algorithm provided by this invention;

[0065] Figure 4 A schematic diagram of the fractal attention mechanism provided by this invention;

[0066] Figure 5 The flowchart of the dense estimation method provided by the present invention. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] This invention discloses a trajectory estimation method for clustered fixed-wing aircraft based on temporal similarity feature information, comprising the following steps:

[0069] S1: Acquire first-person view image data of the cluster of aircraft;

[0070] S2: The small target recognition method based on shallow feature attention performs target recognition on the image data collected in S1 to obtain small target features;

[0071] S3: The dense target post-processing method based on feature matching performs density estimation and dense sample estimation on the small target features obtained in S2 to obtain the filtered small target features;

[0072] S4: The multi-target matching method based on the auction algorithm matches and distinguishes the small target features filtered in S3, thereby achieving matching and distinction of different small target features;

[0073] S5: Using an attention-based target trajectory estimation method, complete the estimation of the feature trajectories of small targets after matching and differentiation in S4.

[0074] Furthermore, the small target recognition method based on shallow feature attention in S2 adopts the FractalAttNet fractal network model, which specifically includes two branches: a feature extraction branch and a filtering branch.

[0075] Specifically, the details of the two branches are as follows:

[0076] The feature extraction branch uses a bottleneck structure to complete feature extraction and utilizes the bottleneck structure to increase the depth of the shallow network.

[0077] The filtering branch uses the HourglassBlock module, which introduces the encoder-decoder structure of the target segmentation and key point recognition network to obtain small-range feature changes. The extracted shallow features are then input into the FractalAttNet fractal network model to complete the extraction of small target features after recognition.

[0078] Furthermore, the FractalAttNet fractal network model is expressed as follows:

[0079]

[0080] Where SE represents the SELayer module, φ(x) represents the convolutional network, Att represents the mini-Hourglass detection module, and wl and w2 represent the feature weights learned by the DropPath method during training of the two branches, respectively.

[0081] Furthermore, S3 specifically includes: using the dense prediction module, after completing target recognition, taking the optimal output extracted from each layer as a reference sample, and using the features extracted from it as similar features, and using the feature matching method to directly complete the dense target segmentation and position estimation in the bottom layer domain.

[0082] Furthermore, the specific content of S4 is: to match the inter-frame feature relationships in the time sequence;

[0083] First, the recognition results between frames are compared, with the recognition result of the previous frame recorded in M ​​and the recognition result of the current frame recorded in N; then, the cost function is used for similarity matching to obtain the final cost matrix.

[0084] The cost function is:

[0085] C ij =||p i -p j ||2*(δS*w s )*lg((0.3δS+1)+1),δS=|S i -S j |·····(2)

[0086] Where p is the predicted center point, S is the predicted bounding box area, δS represents the change in bounding box volume between frames, i represents the frame number of the previous frame, j represents the frame number of the current frame, and w s The area represents the weight in the function, M represents the recognition result set of the previous frame, and N represents the recognition result set of the current frame.

[0087] To ensure the consistency of temporal features, the target extraction results are transformed and projected into an independent matrix position encoding.

[0088] Furthermore, the specific details of S5 are as follows:

[0089] Using the spatial correspondence of analytical geometry, the distance between machines and the size of the target bounding box exhibit a clear inverse relationship, and the calculation formula is as follows:

[0090]

[0091] Among them, s p f is the target pixel size. p is the pixel focal length at the center point of the target, sw is the actual target size, and d is the relative distance.

[0092] Furthermore, S5 also includes:

[0093] A preliminary distance estimation method for monocular cameras is obtained by directly training and converging the relevant parameters using the nonlinear least squares method.

[0094] Establish a negative exponential sample function model:

[0095]

[0096] in, Let x be the negative exponential sample function corresponding to the aircraft distance and bounding box, where a and b are optimization parameters and x is the current bounding box size.

[0097] The objective function obtained from the collected data target value is:

[0098]

[0099] Among them, y i This represents the actual distance value collected from the data, where m is the total number of data points collected.

[0100] Since the collected data are independent and identically distributed, and considering the nonlinear characteristics of the function, the Gauss-Newton method is used to perform a linear transformation on the objective function. The transformation relationship is as follows:

[0101]

[0102]

[0103]

[0104] Where J(x) is the Jacobian matrix of the residual r(x), Let T be the Heather matrix, and let T denote the transpose of the matrix. The first derivative of the function, The second derivative of the objective function;

[0105] Solve the linear equations using Newton's method using formulas (6), (7), and (8):

[0106]

[0107] The corresponding linear relationship is obtained:

[0108] J(x) T J(x)p GN =-J(x) T r(x)·····(10)

[0109] Where, p GN The linear value of the Gauss-Newton method is represented by the linear least squares method, which is used to optimize the objective function.

[0110] By using the Gauss-Newton method to solve for the corresponding values ​​of a and b in the objective function φ(x; a, b), the nonlinear constraint equations are obtained:

[0111]

[0112] A preliminary target distance estimation formula is obtained, and the projection rule is used to complete the preliminary estimation of the target position. Combined with S2, S3, and S4, the target bounding box is accurately extracted to complete the target distance estimation, which is then converted into a single-frame position estimation result.

[0113] Furthermore, the position estimation results are filtered and optimized.

[0114] Convert target location information into a visual radar matrix The numerical coordinates in the matrix represent the polar coordinate projection position of the current target, and the matrix value represents the relative distance of the current target. Using the Transformer structure, a number is added to the value of each aircraft to distinguish different aircraft. The information of each frame is transformed into coded coordinates located in the same space R. The frame number information of consecutive frames is used as the position code input to the Transformer attention network for training to complete the trajectory estimation task of the target.

[0115] In one specific embodiment, the present invention modifies and improves the visual characteristics of swarm aircraft using end-to-end network algorithms. Existing technologies commonly use end-to-end network structures such as... Figure 1 As shown. First, the first-view image data of the swarm aircraft acquired is as follows: Figure 2 As shown in a, for Figure 2 Data labeling was performed to obtain... Figure 2 b. Furthermore, the FractalAttNet feature enhancement module is proposed. The overall structure of this module in the end-to-end network is shown in [see section 1]. Figure 3 See the enhanced module structure. Figure 4 The shallow network is the main component for completing the small target extraction task. In FractalAttNet, the enhancement module designed in this invention includes two branches: a filtering branch and a feature extraction branch. The feature extraction branch is... Figure 4 The lower half directly uses a bottleneck structure module to complete the feature extraction task, and simply increases the depth of the shallow network using the bottleneck structure. The filtering branch is... Figure 4 The upper part contains an HourglassBlock module, which introduces an encoder-decoder structure from the target segmentation and keypoint recognition network to focus more on small-scale feature changes and inputs the extracted shallow features into the subsequent feature enhancement module to complete the small target extraction task. While this lateral fractal network structure can improve the richness of convolutional kernel representation, the effectiveness and quality of features in different branches are difficult to guarantee. Therefore, this invention places a channel-level attention module at the end of each of the two branches, enabling the network to autonomously select and extract appropriate channel features.

[0116] This fractal network module, which we call FractalAttNet in this invention, is expressed as follows:

[0117]

[0118] In this diagram, SE represents the SELayer module, φ(x) represents the convolutional network, and Att represents the mini-Hourglass detection module. w1 and w2 represent the feature weights learned by the DropPath method during training for the two branches, respectively. Finally, the feature layers are re-screened based on the feature scores extracted from the two branches to retain more informative features.

[0119] The FractalAttNet fractal network module is a plug-and-play module with strong network adaptability.

[0120] S3, as a further technical solution, provides a dense target post-processing method based on feature matching as follows:

[0121] After completing the target recognition task, the S2 algorithm outputs the bounding box information of the target on the input image. Typically, end-to-end recognition algorithms directly convert the output results of multiple layers to the input image size during inference, then filter the recognition results using non-maximum suppression based on the confidence of each layer's predictions, retaining the optimal processing result for each anchor point. This invention proposes a novel dense prediction module that, after completing target recognition, uses the optimal output extracted from each layer as a reference sample and its extracted features as similar features. It then utilizes feature matching methods to directly complete the dense target segmentation and location estimation tasks within the bottom-level domain.

[0122] The overall structure of this method is as follows: Figure 5 As shown, firstly, the feature extraction network is trained using a pre-prepared dataset to meet the requirements of target recognition. Then, the dense prediction module is trained step-by-step, while the recognition network is not involved in the training. Next, the Gaussian hotspot method used in the feature point extraction method is employed to generate a prediction feature map, causing the image to converge towards the target's center point, directly predicting the target's location. This directly completes the target's location prediction. Finally, the predicted location is used as the location reference point for the NMS part, and the bounding boxes predicted in the S1 part are processed with each point as the center and the prediction confidence as the ground truth to obtain the optimal prediction box for each point, thus solving the problem of dense target re-identification.

[0123] S4, as a further technical solution, directly matches the inter-frame feature relationships in time based on the auction algorithm, and filters the bounding box width due to the large distance interval between swarm aircraft, thereby improving the overall positioning accuracy.

[0124] First, this invention directly compares the recognition results between frames. The recognition result of the previous frame is recorded in M, and the recognition result of the current frame is recorded in N. Then, the cost function designed in this invention is used for similarity matching to obtain the final cost matrix. The cost function of this invention is:

[0125] C ij =||p i -p j ||2*(δS*w s )*lg((0.3δS+1)+1),δS=|S i -S j |·····(2)

[0126] Where p is the predicted center point, S is the predicted bounding box area, δS represents the change in bounding box volume between frames, i represents the frame number of the previous frame, j represents the frame number of the current frame, and w s This represents the weight of the area in the function, with a default value of 0.7. M represents the recognition result set of the previous frame, and N represents the recognition result set of the current frame.

[0127] Next, the minimum values ​​in each row of the matrix are compared, and it is determined whether there are duplicates in the minimum value sequence. When duplicate values ​​are found, a second screening is performed based on the cost value to complete the target matching task. To ensure the normal updating of the sequence and the effect of lost target recapture, all targets are stored in a message list with the corresponding target number. A preset queue threshold is used to test whether the target exists. If a target has not been updated for a long time, the target is discarded.

[0128] To ensure the consistency of temporal features, this invention transforms all target extraction results and projects them into an independent matrix position encoding, which facilitates the subsequent completion of trajectory estimation tasks.

[0129] As a further technical solution, S5 uses the following target trajectory estimation method based on an attention mechanism:

[0130] In S4, this invention attempts to match and distinguish the identified results, achieving the task of matching and distinguishing different targets within the field of view. When the UAV's position and attitude in the image move parallel to each other and its body axis is aligned with the camera's optical center, the spatial correspondence of analytical geometry can be used to deduce that the distance between the UAVs and the target bounding box size have a clear inverse proportional relationship. The calculation formula is as follows:

[0131]

[0132] Among them, s p f is the target pixel size. p s is the pixel focal length at the center point of the target. w Let d be the actual target size and d be the relative distance. It can be observed that the length of the target bounding box is inversely proportional to the distance between the target and the optical center, and the width of the long side of the target bounding box corresponds to the target's flight distance. This invention directly trains and converges the relevant parameters using a nonlinear least squares method to obtain a preliminary distance estimation method for a monocular camera.

[0133] First, a negative exponential sample function model can be established:

[0134]

[0135] in, Let x be the negative exponential sample function corresponding to the aircraft distance and bounding box, where a and b are optimization parameters and x is the current bounding box size.

[0136] The objective function obtained from the collected data target value is:

[0137]

[0138] Among them, y i This represents the actual distance value collected from the data, where m is the total number of data points collected.

[0139] Since the collected data are independent and identically distributed, and considering the nonlinear characteristics of the function, the Gauss-Newton method is used to perform a linear transformation on the objective function. The transformation relationship is as follows:

[0140]

[0141]

[0142]

[0143] Where J(x) is the Jacobian matrix of the residual r(x), Let T be the Heather matrix, and let T denote the transpose of the matrix. The first derivative of the function, The second derivative of the objective function;

[0144] Solve the linear equations using Newton's method using formulas (6), (7), and (8):

[0145]

[0146] The corresponding linear relationship is obtained:

[0147] J(x) T J(x)p GN =-J(x) T r(x)·····(10)

[0148] Where, p GN The linear value of the Gauss-Newton method is represented by the linear least squares method, which is used to optimize the objective function.

[0149] By using the Gauss-Newton method to solve for the corresponding values ​​of a and b in the objective function φ(x; a, b), the nonlinear constraint equations are obtained:

[0150]

[0151] This leads to a preliminary target distance estimation formula, which is then used to perform a preliminary estimation of the target's position using the projection rule. Combining this with the dense target recognition and differentiation methods proposed in S2, S3, and S4 enables accurate extraction of the target's bounding box. Therefore, the relatively accurate bounding box prediction results can be used to estimate the target's distance, which is then converted into a single-frame position estimation result.

[0152] After obtaining the location prediction result, considering the volatility of predictions made by deep learning methods, the estimated location needs to be filtered and optimized. To address this issue, this invention draws on a 6D pose estimation algorithm based on time information commonly used in the field of autonomous driving. First, the target location information is converted into a visual radar matrix. In the matrix, the numerical coordinates represent the polar coordinate projection position of the current target, and the matrix values ​​represent the relative distance to the current target. Then, to better express the relative constraints between the various aircraft, a Transformer structure is introduced. A unique identifier is added to each aircraft's location to distinguish it from other aircraft.

[0153] Using this method, the present invention transforms the information of each frame into coded coordinates located in the same spatial space R. Then, the frame number information of consecutive frames is used as the position code input to the Transformer attention network for training, completing the trajectory estimation task for the target.

[0154] It can be observed that the final estimated trajectory of the present invention is basically consistent with the actual flight trajectory, demonstrating good trajectory estimation performance.

[0155] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A trajectory estimation method for a cluster of fixed-wing aircraft based on timing similarity feature information, characterized in that, The method comprises the following steps: S1: collecting first-view image data of a cluster aircraft; S2: performing target recognition on the image data collected in S1 based on a shallow feature attention small target recognition method to obtain small target features; S3: performing density estimation and dense sample estimation on the small target features obtained in S2 based on a feature matching dense target post-processing method to obtain screened small target features; S4: performing matching and distinguishing on the screened small target features in S3 based on a multi-target matching method based on an auction algorithm to realize matching and distinguishing of different small target features; S5: using a target trajectory estimation method based on an attention mechanism to complete trajectory estimation of the small target features matched and distinguished in S4; The shallow feature attention small target recognition method in S2 adopts a FractalAttNet fractal network model, which specifically includes two branches: a feature extraction branch and a filtering branch. The FractalAttNet fractal network model is expressed by a formula as follows: wherein SE represents an SE Layer module, represents a convolutional network, Att represents a mini-Hourglass detection module, w 1 and w 2 respectively represent feature weights learned by the two branches using the DropPath method during training. S3 specifically includes: using a dense prediction module to, after completing target recognition, take the optimal output extracted from each layer as a reference sample, and taking the extracted features as similar features, and using a feature matching method to directly complete dense target division and position estimation in the bottom domain. The specific content of S4 is to match the inter-frame feature relationship in time sequence. First, the recognition results between frames are compared, the recognition result of the previous frame is recorded as the recognition result of the current frame is recorded as ; then, similarity matching is performed using a cost function to obtain a final cost matrix; The cost function is as follows: wherein, is a predicted center point, is a predicted bounding box area, represents a change in the volume of the inter-frame bounding box, represents a frame number of a previous frame, j represents a frame number of a current frame, represents a weight of the area in the function, represents a recognition result set of the previous frame, represents a recognition result set of the current frame; In order to ensure the consistency of the time sequence features, the target extraction result is converted and projected into an independent matrix position coding.

2. The cluster fixed-wing aircraft trajectory estimation method based on timing similarity feature information according to claim 1, characterized in that, The specific content of S5 is as follows: Using the spatial correspondence relationship of analytic geometry, the inter-aircraft distance and the target bounding box size have a clear inverse proportional relationship, and the calculation formula is as follows: wherein, is the target pixel size, is the pixel focal length at the target center point, is the actual target size, is the relative distance.

3. The cluster fixed-wing aircraft trajectory estimation method based on timing similarity feature information according to claim 1, characterized in that, S5 further includes: Through a nonlinear least squares method, the related parameters are directly trained and converged to obtain a preliminary distance estimation method of a monocular camera; A negative exponential sample function model is established: wherein, is a negative exponential sample function corresponding to the aircraft distance and the bounding box, is an optimization parameter, is the current bounding box size; Using the collected data target value, the target function is obtained as follows: wherein, y i represents the real distance value of data collection, m is the total number of data collected; Since the collected data is independently and identically distributed, considering the nonlinear characteristics of the function, a Gauss-Newton method is used to linearly convert the target function, and the conversion relationship is as follows: wherein is the Jacobian matrix of the residual is the Hessian matrix T denotes the transpose of a matrix is the first order derivative of the function is the second order derivative of the objective function​ The Gauss-Newton linear equation is solved by using formulas (6), (7), and (8): The corresponding linear relationship is obtained as follows: wherein, represents a linear value of the Gauss-Newton method, and the objective function is optimized by a linear least square method; Solve the objective function using the Gauss-Newton method In corresponding values of the parameters to obtain a nonlinear constraint equation: The preliminary target distance estimation formula is obtained, the projection method is used to complete the preliminary estimation of the target position, the target bounding box is accurately extracted by combining S2, S3, and S4, and the distance estimation of the target is completed, and then the position estimation result of a single frame is obtained.

4. The cluster fixed-wing aircraft trajectory estimation method based on timing similarity feature information according to claim 3, characterized in that, The position estimation result is filtered and optimized. Converting target position information into visual radar matrix The numerical coordinates in the matrix represent the polar coordinate projection position of the current target, and the matrix values represent the relative distance of the current target. A number value is added to each aircraft value using the Transformer structure to distinguish different aircraft. The information of each frame is converted into encoded coordinates located in the same space R. The frame number information of the continuous frames is input into the Transformer attention network for training to complete the trajectory estimation task of the target.

Images