An aerial cone target detection and positioning method

Abstract

The application discloses an air cone sleeve target detection and positioning method and belongs to the technical field of target tracking and positioning; the application provides an air cone sleeve target tracking and positioning method; firstly, infrared target image acquisition is completed by installing near-infrared LED lamp beads on the cone sleeve and installing an infrared filter on a camera lens. Then, an improved YOLOv11 detection algorithm is provided, the detection capability of YOLO for small targets is improved, and the Bytetrack strategy is adopted to realize infrared target tracking. Finally, an improved pose estimation method is used to improve the positioning precision of the cone sleeve position.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target tracking and positioning, and particularly relates to an air cone target detection and positioning method. BACKGROUND

[0002] Motion target tracking and positioning has always been a challenging task and is also a hot research direction at present.

[0003] In terms of target tracking, early target tracking algorithms mainly track targets according to target modeling or target features, and the main methods include:

[0004] 1) Target model-based modeling method. The target appearance model is modeled, and then the target is found in the subsequent frame.

[0005] 2) Search-based method. With the in-depth research, it is found that the target model-based modeling method processes the whole picture, and the real-time performance is poor.

[0006] Recently, with the wide application of deep learning methods, people have begun to consider applying them to target tracking. At present, the application of deep learning in the field of target tracking can be divided into two categories. One is to combine the correlation filtering method, and to use deep learning as a feature extraction method and to use correlation filtering as a tracking framework for target tracking. The other is a completely neural network-based target tracking method. Among them, the YOLO series algorithm is highly concerned. Although the YOLO series algorithm performs well in real-time and many other aspects, there are problems of missed detection and false detection in small target detection. Target detection algorithm. YOLO (You Only Look Once) is an end-to-end target detection algorithm.

[0007] In terms of pose positioning, the traditional target positioning technology mainly uses radar, laser, infrared and other sensors for ranging to achieve target positioning. However, these sensors have the following shortcomings: on the one hand, they are easily affected by other sensors carried by the machine itself, resulting in inaccurate ranging; on the other hand, their biggest disadvantage is that they cannot identify target objects. Visual positioning methods can be divided into monocular vision-based positioning, binocular vision-based positioning and omnidirectional vision-based positioning. The monocular vision-based positioning method can be divided into single-image-based, two-image-based or more than two-image-based monocular vision positioning according to the number of images used for positioning. The single-image-based positioning determines the pose relationship between the world coordinate system and the camera coordinate system according to the description of the known target features (such as points, lines, curves, planes, quadrilaterals, etc.) in the world coordinate system, the projection of these features on the image, and some additional feature parameters, to position the target. SUMMARY

[0008] The purpose of the present application is to provide an air cone sleeve target detection and positioning method to solve the above problems, which completes infrared target image acquisition by installing near-infrared LED lamp beads on the cone sleeve and installing an infrared filter on the camera lens. Then, an improved YOLOv11 detection algorithm is proposed to improve the detection ability of small targets by YOLO, and the Bytetrack strategy is used to realize infrared target tracking. Finally, an improved pose estimation method is used to improve the positioning accuracy of the cone sleeve position.

[0009] The present application is realized by the following scheme:

[0010] An air cone sleeve target detection and positioning method, comprising the following steps:

[0011] Step S1, model training, using the near-infrared annotation dataset of the air cone sleeve to train the yolo detection model;

[0012] Step S2, calibrate the camera, solve the internal and external parameter matrices and distortion parameters of the camera, and optimize the internal parameters of the camera using the CPSO algorithm;

[0013] Step S3, use homomorphic filtering technology to pre-process the real-time collected images;

[0014] Step S4, use the trained yolo detection model and Bytetrack strategy to complete real-time target detection and tracking of the air cone sleeve;

[0015] Step S5, according to the boundary box detected by the yolo detection model, extract the region of interest and perform filtering and binarization processing to convert it into a binary image;

[0016] Step S6, perform Hough transform on the binary image spot area, fit the ellipse formed by the spot, and solve the ellipse center coordinates to obtain the coordinates of the cone sleeve center position in the image;

[0017] Step S7, use the calibrated camera parameters to calculate the relative position of the target in the world coordinate system and the camera, and then calculate the position offset of the fixed point of the unmanned aerial vehicle and the center of the cone sleeve according to the position offset of the camera and the fixed point of the unmanned aerial vehicle.

[0018] In step S1, the specific steps of infrared target image acquisition are:

[0019] The 8 infrared LED light sources installed on the cone sleeve are used as target recognition objects, and the 8 infrared LED light sources are evenly spaced around the circumferential position of the cone sleeve. The camera lens is installed with an infrared filter, and the sampling is performed through the camera.

[0020] In step S2, Zhang Zhengyou calibration method is used for camera calibration, and the camera internal parameters are optimized; specifically,

[0021] According to the linear relationship between the image coordinate system, the pixel coordinate system, the camera coordinate system and the world coordinate system, a mathematical model is constructed;

[0022] The chaos particle swarm optimization algorithm is used in the optimization process of the camera's internal and external parameters, so that the camera parameter solution converges to the global optimal solution with a high probability, and the calibration parameters with higher precision are obtained; the chaos particle swarm optimization algorithm selects the Logistics equation as the chaos system.

[0023] In step S2, the mathematical model constructed is:

[0024]

[0025] Wherein: u, v are the horizontal and vertical coordinate values of the image coordinate system, S is the distance from the camera image plane to the object along the optical axis, dx and dy are the pixel horizontal and vertical dimensions, u0 and v0 are the center positions of the image plane, f is the distance of the camera, R is the rotation matrix of the calibration object, t is the translation matrix, kx and ky are the magnification coefficients of the imaging coordinates on the normalized imaging plane to the image coordinates, X, Y and Z are the positions of the feature points in the world coordinate system;

[0026] The mathematical model is simplified to obtain the following formula:

[0027]

[0028] Wherein: S is a scaling factor, M is the object plane coordinate, A1 is the internal parameter matrix of the camera, A2 is the external parameter matrix of the camera, H is the homography matrix, which contains the internal and external parameters of the camera;

[0029] The required feature points are determined by the formula, and the chaos particle swarm optimization algorithm is used in the optimization process of the camera's internal and external parameters;

[0030] Specifically, the following objective function is optimized by using the CPSO algorithm:

[0031]

[0032] The optimal solution of the camera internal parameter is obtained, wherein: The world coordinate point M is obtained by projecting the camera imaging model; m ij The pixel coordinate point obtained by corner detection is:

[0033] The evolutionary strategy of inertia weight is adopted: the nonlinearity is reduced, and the specific formula is:

[0034] :

[0035] In the formula, iter maxw is a maximum iteration number, w k w is an inertia weight at the kth iteration, w min w is a value range of w max w is a value range of w k

[0036] Step S3 is specifically: before target detection and tracking, the image will be preprocessed by homomorphic filtering technology on the original image, and the homomorphic filtering is an image enhancement and repair technology, which separates the illumination component and the reflection component of the image, and recombines them after separate processing, and the specific steps include: logarithmic transformation, Fourier transformation, high-pass filtering, inverse Fourier transformation and exponential transformation.

[0037] In step S4, the trained yolo detection model and the Bytetrack strategy are used to complete the real-time target detection and tracking of the air cone sleeve; specifically, the following steps are included:

[0038] Step S41: improvement of yolo algorithm, all ReLu activation functions in YOLOv11 backbone network are replaced with SiLU function; an ECANet module is introduced in the backbone network of YOLOv11;

[0039] Step S42: target tracking based on yolo-Bytetrack, the detector of YOLOv11 will use Bytetrack to complete tracking after completing the target detection of the cone sleeve, Bytetrack is a tracking algorithm based on target detection, which uses Kalman filter to predict the bounding box, and uses Hungarian algorithm to match the target detection box and the predicted box; in addition, two associations are performed through high-confidence detection boxes and low-confidence detection boxes.

[0040] In step S41, for a given input feature, ECANet first performs global average pooling operation on each channel and keeps the channel dimension unchanged; then, the processed data is transmitted to 1x1 convolution for information fusion; then, the data is activated by Sigmoid function and multiplied with the original feature, so as to fully fuse the global feature.

[0041] In step S5, it specifically includes the following steps:

[0042] Step S51: region of interest extraction, assuming that the position of the bounding box detected by yolo is (x1, y2, h, w), where x, y, h and w are the center horizontal coordinate, the center horizontal coordinate, the height and the width of the detection box, respectively, when cropping, to ensure that the detection target is in the cropped region, a region with a certain width d will be expanded outward when cropping, then the bounding box information of the region of interest is (x1, y2, h+d, w+d);

[0043] ​Step S52: calculate the front elliptical center coordinates of the cone sleeve, and the algorithm filters and binarizes the region of interest after cutting.

[0044] In step S6, it specifically includes the following steps:

[0045] Step S61, solve the elliptical center coordinates: use Hough transform to fit the light spot in the region with an ellipse, and solve the center coordinates (x2, y2) of the ellipse in the o'-x'y' coordinate system. The information of the detection bounding box in the O-xy coordinate system is (x1, y2, h, w), and the coordinates of the o' point in the O-xy coordinate system are (x1-0.5h-d, y1-0.5h-d). The center coordinates of the ellipse in the O-xy coordinate system are (x1+x2-0.5h-d, y1+y2-0.5h-d);

[0046] Step S62, solve the coordinates of the cone sleeve center on the image, the linear relationship between the image coordinate system, the pixel coordinate system, the camera coordinate system and the world coordinate system, and the calibrated camera internal and external parameters, the homography matrix H value, and use the related formula to solve the coordinates of the cone sleeve center in the real world relative to the camera:

[0047] .

[0048] In step S7, it specifically includes the following steps:

[0049] Since there is a coordinate deviation (Δ x , Δ y , Δ z ) between the camera and the fixed point of the unmanned aerial vehicle in the real world, the deviation value is obtained by real measurement, and the coordinate deviation needs to be considered when solving the relative position of the cone sleeve and the fixed point of the unmanned aerial vehicle in the real world. The coordinates of the cone sleeve relative to the unmanned aerial vehicle are W +(Δ x , Δ y , Δ z ).

[0050] As described above, due to the adoption of the above technical scheme, the present application has the following beneficial effects:

[0051] 1. The present application provides an improved YOLOv11 detection algorithm, which improves the detection ability of YOLO for infrared small targets, and uses the Bytetrack strategy for tracking. Then, an improved pose estimation method is used to measure the target position, thereby improving the positioning accuracy of the target. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is a technical flow chart of an aerial cone sleeve target tracking and positioning method; BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 2 Search flow chart for optimizing camera intrinsic parameters by CPSO algorithm;

[0054] Figure 3 Front view installation of infrared LED flat effect drawing of the cone sleeve proposed in the application;

[0055] Figure 4 3D effect drawing of the cone sleeve;

[0056] Figure 5 Front view effect drawing of the cone sleeve with infrared LED under camera infrared imaging;

[0057] Figure 6 YOLOv11 backbone network improvement drawing;

[0058] Figure 7 Yolo-Bytetrack target tracking algorithm flow chart;

[0059] Figure 8 Region of interest extraction method drawing;

[0060] Figure 9 Ellipse fitting effect drawing;

[0061] Figure 10 Region of interest-camera coordinate transformation drawing. DETAILED DESCRIPTION

[0062] All features disclosed in this specification, and / or the steps of any method or process specified in this specification, can be combined in any combination, except combinations where at least some of the features and / or steps are mutually exclusive.

[0063] Any feature disclosed in this specification, unless stated otherwise, can be replaced by any other feature serving the same, or a similar, purpose.

[0064] In the description of the application, it should be understood that the orientation or positional relationship indicated by the terms "upper", "lower", "left", "right", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the purpose of facilitating the description of the application and simplifying the description, and does not indicate or imply that the device or element referred to must have a predetermined orientation, be constructed and operated in a predetermined orientation, and therefore cannot be understood as a limitation on the application.

[0065] In addition, the terms "first", "second", etc. are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined with "first", "second", etc. can explicitly or implicitly include one or more of the features.

[0066] Example 1

[0067] The present invention provides a technical solution:

[0068] like Figure 1 As shown, a method for detecting and locating an aerial drogue target includes the following steps:

[0069] Step S1, model training, using the near-infrared annotated dataset of the aerial drogue to train the YOLO detection model;

[0070] In step S2, the camera is calibrated using the Zhang Zhengyou calibration method to solve the camera's intrinsic and extrinsic parameter matrix and distortion parameters, and the camera's intrinsic parameters are optimized using the CPSO algorithm (chaotic particle swarm optimization algorithm).

[0071] Step S3, using homomorphic filtering technology to perform image preprocessing on the real-time collected image to solve the problem of uneven illumination of the image;

[0072] Step S4, using the trained YOLO detection model and Bytetrack strategy to complete real-time target detection and tracking of the aerial drogue;

[0073] Step S5: extract the region of interest (ROI) based on the bounding box detected by the Yolo detection model, filter and binarize it, and convert it into a binary image;

[0074] Step S6, performing Hough transform on the spot area of ​​the binary image, fitting the ellipse formed by the spot, solving the coordinates of the center of the ellipse to obtain the coordinates of the center position of the cone sleeve on the image;

[0075] In step S7, the relative position of the target and the camera in the world coordinate system is calculated using the calibrated camera parameters. Then, based on the position offset between the camera and a fixed point of the drone, the position offset between a fixed point of the drone and the center of the drogue is calculated.

[0076] In step S1, before model training, it is necessary to first prepare an infrared target dataset. The specific steps are as follows:

[0077] The 8 infrared LED light sources installed on the cone sleeve are used as target identification objects. The 8 infrared LED light sources are evenly spaced around the circumference of the cone sleeve. Its front plane structure is as follows Figure 3 As shown, the 3D structure is Figure 4 As shown; the camera lens is equipped with an infrared filter, and its sampling effect is shown in the figure Figure 5 As shown; the YOLO model is based on Figure 5 The sample data set is used to train the detection model;

[0078] In step S2, the present application adopts Zhang Zhengyou calibration method in camera calibration, and optimizes the camera internal parameters, which specifically includes the following steps:

[0079] According to the imaging relationship, the linear relationship between the image coordinate system, the pixel coordinate system, the camera coordinate system and the world coordinate system, the following mathematical model is constructed:

[0080] (1)

[0081] Wherein: u, v are the horizontal and vertical coordinate values of the image coordinate system, S is the distance from the camera image plane to the object along the optical axis, dx, dy are the horizontal and vertical dimensions of the pixel, u0, v0 are the center positions of the image plane, f is the distance of the camera, R is the rotation matrix of the calibration object, t is the translation matrix, kx, ky respectively represent the magnification coefficients of the imaging coordinates on the normalized imaging plane to the image coordinates, X, Y and Z are the positions of the feature points in the world coordinate system;

[0082] The formula (1) is simplified as:

[0083] (2)

[0084] Wherein: S is the scaling factor scalar, M is the object plane coordinate, A1 is the internal parameter matrix of the camera, A2 is the external parameter matrix of the camera (rotation and translation parameters of the calibration board, [r1, r2, t]), H is the homography matrix, which contains the internal and external parameters of the camera.

[0085] As can be seen from formula (2), the matrix equation can be converted into two linear matrices, so as to obtain the 9 unknown factors in the homography matrix H, at least 5 feature points are needed in space, and with the increase of the feature points, the LM method can obtain more accurate internal parameter values; (Levenberg-Marquardt method (LM method) is a nonlinear optimization algorithm combining gradient descent method and Gauss-Newton method, mainly used to solve least square problem, widely used in machine learning and engineering field).

[0086] In the actual application process, due to manufacturing, installation and other factors, the camera will inevitably have various distortions; in order to describe the law of distortion imaging, a nonlinear imaging model is introduced; therefore, there is deviation in the actual lens imaging, and these deviation parts need to be considered in the process of solving the internal and external parameters, the coordinates of the actual image point are (u d , v d ), the ideal coordinates are (u u , v u ), usually only considering the radial distortion can achieve high precision and meet the application requirements, the corresponding relationship is shown in formula (3):

[0087] (3)

[0088] In the formula: k1, k2, k3 are distortion parameters, which represent 2nd, 4th and 6th distortion parameters in horizontal and vertical directions.

[0089] In the above Zhang Zhengyou camera calibration, the initial value of the camera is obtained by assuming that the camera is a linear model, and the initial value of the homography matrix H is solved by optimizing the initial value by the LM method; but the Zhang Zhengyou calibration method is based on the linear model of the camera to estimate the distortion parameters, and the traditional optimization strategy such as the LM method is easy to fall into local optimum, and the calibration accuracy will be reduced; in order to improve the accuracy of the camera calibration parameters, the chaos particle swarm optimization algorithm (CPSO) is used in the optimization process of the internal and external parameters of the camera calibration, so that the camera parameter solution can converge to the global optimal solution with high probability, and the calibration parameters with higher accuracy are obtained.

[0090] The following objective function is optimized using the CPSO algorithm:

[0091] (4)

[0092] The optimal solution of the camera internal parameter is obtained, wherein: The world coordinate point M is obtained by projecting the camera imaging model; m ij The pixel coordinate point is obtained by detecting the corner point.

[0093] The CPSO is to map the chaotic variable from the chaos space to the solution space, and then perform random search on the chaotic variable. The chaos optimization algorithm has the characteristics of not sensitive to initial value, easy to jump out of local minimum, fast search speed, high calculation precision and global asymptotic convergence.

[0094] The evolution strategy of the inertia weight is nonlinearly reduced as formula (5):

[0095] (5)

[0096] In the formula, iter max Is the maximum iteration number, w k Is the inertia weight at the kth iteration, w min , w max Is the value range of w k .

[0097] In addition, the CPSO algorithm selects the Logistics equation as the chaotic system, and the calculation amount of the equation is smaller than that of other chaotic systems, so the equation is used to construct the chaotic sequence; the search process of the chaos particle swarm optimization algorithm (CPOS) is shown in Figure 2 .

[0098] In step S3, the real-time collected image is preprocessed by using the homomorphic filtering technology, specifically, before target detection and tracking, the image is preprocessed by using the homomorphic filtering technology on the original image, the homomorphic filtering is an image enhancement and repair technology, which separates the illumination component and the reflection component of the image, and recombines them after separate processing, the specific steps include logarithmic transformation, Fourier transformation, high-pass filtering, inverse Fourier transformation and exponential transformation, and the technology has good effect on the problem of uneven illumination of the processed image;

[0099] In step S4, the trained yolo detection model and the Bytetrack strategy are used to complete the real-time target detection and tracking of the air cone sleeve; Specifically, the following steps are included:

[0100] Step S41: improvement of yolo algorithm, in order to identify and track the cone sleeve, a plurality of training data sets with boundary box labels need to be made, the present application considers that the distance between the cone sleeve and the camera is far away in the air, the cone sleeve occupies a small proportion in the sampling image, and there is small target identification. Therefore, the present application proposes an improved detection and tracking algorithm based on YOLOv11, which replaces all ReLu activation functions in the YOLOv11 backbone network with SiLU functions. SiLU function can effectively perform nonlinear transformation on the input of neural network, and is differentiable in the entire definition domain. SiLU function is not monotonically increasing, but reaches a global minimum value of-0.28 at x≈-1.28, which can play a role of implicit regularization and suppress excessive weights.

[0101] Secondly, the algorithm also integrates the attention mechanism module, the application of the attention module can significantly improve the target detection performance. The existing attention mechanisms include CBAM, SOCA, SENet, ECANet, etc. Adding these different attention mechanisms has been confirmed in related research that they can improve the performance of the original YOLO network structure; ECANet is very suitable for CNN, which can effectively enhance the cross-channel interaction information while avoiding model complexity and information loss problems; in order to improve the detection performance of small targets by the model, the ECANet module is introduced into the backbone network of YOLOv11, as shown in Figure 6

[0102] For the given input features, ECANet first performs global average pooling operation on each channel and keeps the channel dimension unchanged; then, the processed data is input into a 1x1 convolution for information fusion; then, the data is activated by a Sigmoid function and multiplied with the original features to fully fuse the global features;

[0103] ​Step S42: yolo-Bytetrack-based target tracking. After completing the target detection of the cone sleeve, the detector of YOLOv11 uses Bytetrack to complete tracking. Bytetrack is a tracking algorithm based on target detection, which uses Kalman filtering to predict the bounding box and uses the Hungarian algorithm to match the target detection box and the predicted box. In addition, two associations are performed through high-confidence detection boxes and low-confidence detection boxes, which can achieve good tracking effect. The process of the yolo-Bytetrack-based target tracking algorithm is shown in FIG. 8. Figure 7

[0104] In step S5, the main idea of this step is: in order to more accurately calculate the pose of the front center of the cone sleeve and the camera, after completing the target bounding box detection, the algorithm extracts and crops the region of interest, and then uses Hough transformation to fit the ellipse in the cropped region of interest, and finds the center coordinates of the ellipse. It specifically includes the following steps:

[0105] Step S51: region of interest extraction. Assuming that the position of the bounding box detected by yolo is (x1, y2, h, w), where x, y, h, and w are the center horizontal coordinate, center horizontal coordinate, height, and width of the detection box, respectively. When cropping, in order to ensure that the detection target is within the cropped region, a region with a certain width d is expanded outward according to the bounding box, as shown in FIG. 7. Therefore, the bounding box information of the region of interest is (x1, y2, h+d, w+d). Figure 8

[0106] Step S52: calculate the center coordinates of the front ellipse of the cone sleeve. The algorithm filters and binarizes the cropped region of interest.

[0107] In step S6, it specifically includes the following steps:

[0108] Step S61: solve the center coordinates of the ellipse. Hough transformation is used to fit the ellipse in the region, and the center coordinates of the ellipse in the o'-x'y' coordinate system are solved (x2, y2). As known from the foregoing, the information of the detection bounding box in the O-xy coordinate system is (x1, y2, h, w), and the coordinates of the o' point in the O-xy coordinate system are (x1-0.5h-d, y1-0.5h-d). Therefore, the center coordinates of the ellipse in the O-xy coordinate system (image coordinate system) are (x1+x2-0.5h-d, y1+y2-0.5h-d).

[0109] ​​Step S62, solve the coordinates of the center of the cone sleeve on the image, according to the linear relationship between the image coordinate system, the pixel coordinate system, the camera coordinate system and the world coordinate system, and the camera internal and external parameters solved by calibration, the homography matrix H value, the formula (2) can be used to solve the coordinates of the center of the cone sleeve in the real world relative to the camera:

[0110] (7)

[0111] In step S7, specifically, due to the existence of coordinate deviation (Δ x , Δ y , Δ z ) between the camera and the fixed point of the unmanned aerial vehicle in the real world, the deviation value can be obtained by real measurement, therefore, when solving the relative position of the cone sleeve and the fixed point of the unmanned aerial vehicle in the real world, the coordinate deviation needs to be considered, and the coordinates of the cone sleeve relative to the unmanned aerial vehicle are W +(Δ x , Δ y , Δ z ).

[0112] The above only describes the preferred embodiments of the present application and is not intended to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for detecting and locating an aerial drogue target, characterized by: The following steps are involved: Step S1, infrared target image acquisition, and use the near-infrared annotated dataset of the aerial drogue to train the YOLO detection model; Step S2, calibrating the camera, solving the camera's intrinsic and extrinsic parameter matrix and distortion parameters, and optimizing the camera's intrinsic parameters using the CPSO algorithm; Step S3, performing image preprocessing on the real-time collected images using homomorphic filtering technology; Step S4, using the trained YOLO detection model and Bytetrack strategy to complete real-time target detection and tracking of the aerial drogue; In step S4, the trained YOLO detection model and Bytetrack strategy are used to complete the real-time target detection and tracking of the aerial drogue. Specifically, the following steps are included: Step S41: Improve the YOLO algorithm by replacing all ReLu activation functions in the YOLOv11 backbone network with SiLU functions; introduce the ECANet module into the YOLOv11 backbone network; In step S41, ECANet first performs a global average pooling operation on each channel for a given input feature while keeping the channel dimension unchanged. Then, the processed data is passed to a 1×1 convolution for information fusion. Subsequently, the data is activated by a Sigmoid function and multiplied with the original feature to fully fuse the global features. Step S42: Target tracking based on YOLO-Bytetrack. After completing the target detection of the cone sleeve, the YOLOv11 detector will use Bytetrack to complete the tracking. Bytetrack is a tracking algorithm based on target detection. It uses Kalman filtering to predict the bounding box and uses the Hungarian algorithm to match the target detection box and the predicted box. In addition, it performs two associations through the high-confidence detection box and the low-confidence detection box. Step S5: extract the region of interest based on the bounding box detected by the Yolo detection model, filter it, binarize it, and convert it into a binary image; Step S6, performing Hough transform on the spot area of ​​the binary image, fitting the ellipse formed by the spot, solving the coordinates of the center of the ellipse to obtain the coordinates of the center position of the cone sleeve on the image; In step S7, the relative position of the target and the camera in the world coordinate system is calculated using the calibrated camera parameters. Then, based on the position offset between the camera and a fixed point of the drone, the position offset between a fixed point of the drone and the center of the drogue is calculated.

2. The method for detecting and locating an aerial drogue target according to claim 1, wherein: In step S1, the specific steps of infrared target image acquisition are: The eight infrared LED light sources installed on the cone sleeve serve as target identification objects. The eight infrared LED light sources are evenly spaced around the circumference of the cone sleeve. The camera lens is equipped with an infrared filter, and filtered sampling is performed through the camera.

3. The method for detecting and locating an aerial drogue target according to claim 2, wherein: In step S2, the camera is calibrated using the Zhang Zhengyou calibration method and the camera internal parameters are optimized; specifically, Construct a mathematical model based on the imaging relationship, the linear relationship between the image coordinate system, the pixel coordinate system, the camera coordinate system and the world coordinate system; The chaotic particle swarm optimization algorithm is used in the optimization process of internal and external parameters of camera calibration, so that the camera parameter solution converges to the global optimal solution with a high probability, and obtains calibration parameters with higher accuracy; the chaotic particle swarm optimization algorithm uses the Logistics equation as the chaotic system.

4. The method for detecting and locating an aerial drogue target according to claim 3, wherein: In step S2, the mathematical model constructed is: Where: u, v are the horizontal and vertical coordinates of the image coordinate system, S is the distance from the camera image plane to the object along the optical axis, dx, dy are the horizontal and vertical dimensions of the pixel, u0, v0 are the center positions of the image plane, f is the distance to the camera, R is the rotation matrix of the calibration object, t is the translation matrix, k x , k y They represent the magnification coefficients from the imaging coordinates to the image coordinates on the focal length normalized imaging plane, respectively. X, Y, and Z are the positions of the feature points in the world coordinate system. The following formula is obtained by simplifying the mathematical model: Where: M is the object plane coordinate, A1 is the camera's intrinsic parameter matrix, A2 is the camera's extrinsic parameter matrix, and H is the homography matrix, which contains both the camera's intrinsic and extrinsic parameters. The required feature points are determined through the formula, and the chaotic particle swarm optimization algorithm is used in the process of optimizing the internal and external parameters of the camera calibration; Specifically: Use the CPSO algorithm to optimize the following objective function: Get the optimal solution of the camera's internal parameters, where: The object plane coordinate point M is projected through the camera imaging model; m ij is the pixel coordinate point obtained by corner point detection: The evolutionary strategy of inertia weight is adopted to reduce nonlinearity. The specific formula is: ： In the formula, iter max is the maximum number of iterations, w k is the inertia weight at the kth iteration, w min 、w max w k The value range of .

5. The method for detecting and locating an aerial drogue target according to claim 4, wherein: Step S3 is specifically as follows: before target detection and tracking, the image will be preprocessed using homomorphic filtering technology. Homomorphic filtering is a technology used for image enhancement and restoration. It separates the illumination component and the reflection component of the image, processes them separately, and then recombine them. The specific steps include: logarithmic transformation, Fourier transform, high-pass filtering, inverse Fourier transform, and exponential transformation.

6. The method for detecting and locating an aerial drogue target according to claim 5, wherein: In step S5, it specifically includes the following steps: Step S51: Extracting the region of interest. Assume that the bounding box position detected by Yolo is (x1, y2, h, w), where x1, y2, h, and w are the center horizontal coordinate, center vertical coordinate, height, and width of the detection box, respectively. When cropping, to ensure that the detection target is within the cropping area, the bounding box will be expanded outward by a certain width d. The bounding box information of the region of interest is (x1, y2, h+d, w+d); Step S52: Calculate the center coordinates of the front ellipse of the cone sleeve, and the algorithm performs filtering and binarization processing on the cropped region of interest.

7. The method for detecting and locating an aerial drogue target according to claim 6, wherein: In step S6, it specifically includes the following steps: Step S61, solve the center coordinates of the ellipse: Use the Hough transform to fit the light spot in the area to an ellipse, and solve for the center coordinates of the ellipse in the o'-x'y' coordinate system (x2, y2). The detection bounding box information in the O-xy coordinate system is (x1, y2, h, w). The coordinates of point o' in the O-xy coordinate system are (x1-0.5hd, y1-0.5hd). The center coordinates of the ellipse in the O-xy coordinate system are solved as (x1+x2-0.5hd, y1+y2-0.5hd). Step S62: Calculate the coordinates of the center of the cone sleeve on the image, the linear relationship between the image coordinate system, the pixel coordinate system, the camera coordinate system, and the world coordinate system, and calibrate the solved camera internal and external parameters and the homography matrix H value. Use the relevant formula to calculate the coordinates of the cone sleeve center relative to the camera in the real world: 。 8. The method for detecting and locating an aerial drogue target according to claim 7, wherein: In step S7, it specifically includes the following steps: Since there is a coordinate deviation between the camera and the fixed point of the drone in the real world (Δ x , Δ y , Δ z ), specifically, the deviation value is obtained through real-world measurement. The coordinate deviation must be considered when solving the real-world relative position of the drogue and a fixed point of the drone. The coordinate of the drogue relative to the drone is W +(Δ x , Δ y , Δ z ).

Images