A method for fast generation of ground object classification images for UAV remote sensing image sequences

Abstract

The present invention discloses a method for quickly generating a ground object classification image for a UAV remote sensing image sequence, and belongs to the field of UAV remote sensing technology. The present invention first generates a ground object classification map that completely covers the survey area through visible light, multi-spectral and other data that are easy to classify ground objects; then the UAV remote sensing image sequence to be classified is spliced ​​into a mosaic, and the classification map is registered and resampled with it, and the calibrated camera space position, posture and other parameters are derived at the same time; for a certain frame image in the image sequence to be classified, the scale and the offset of the image center relative to the photography center are calculated in combination with the camera's internal orientation elements and the spatial resolution of the mosaic, so as to determine its row and column number in the classification map. Finally, the corresponding classification sub-map can be extracted from the classification map in combination with the size of the image frame and the film rotation angle. The present invention can quickly and accurately obtain the ground object classification image sequence corresponding to the remote sensing image sequence, and has good scalability and practicality.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle remote sensing technology, and in particular to a method for quickly generating ground object classification images for unmanned aerial vehicle remote sensing image sequences. Background Art

[0002] UAV remote sensing image classification is of great significance in the fields of agriculture, environmental monitoring and ecological protection. Through high-resolution images, drones can provide accurate land use information to help personnel achieve precision agricultural management; in environmental monitoring, they can monitor natural disasters, water quality changes, etc., and improve emergency response efficiency; in ecology, through classification analysis, it helps to protect wild animals and plants and their habitats and promote sustainable ecological balance.

[0003] Commonly used UAV remote sensing image classification methods are mainly divided into supervised classification and unsupervised classification. Supervised classification uses sample data of known categories for training, learns the features of samples through algorithms, and then applies these learned features to the entire image for classification; its advantages are high accuracy, applicable to scenes of known categories, and strong interpretability of results; but this method requires labeled training data and is sensitive to data quality and label accuracy. Unsupervised classification does not use prior knowledge, and divides images into different categories based on the intrinsic statistical characteristics and structure of the image; its advantages are that it does not require a large amount of labeled training data, is applicable to scenes of unknown categories, and can automatically identify potential patterns in the data; but the obtained classification results are less interpretable and are more sensitive to image quality and scene complexity. Previous methods usually perform related operations on ortho-mosaics, and direct application to UAV remote sensing image sequences requires a large amount of data annotation and manual adjustment to obtain a more reliable classified image sequence. This deficiency limits the (near) real-time and high-accuracy interpretation of UAV remote sensing data, thereby restricting its application prospects in precision agriculture, ecological monitoring, etc. Summary of the invention

[0004] The purpose of the invention is to provide a method for quickly generating ground object classification images for UAV remote sensing image sequences by adopting the idea of ​​"image backtracking" in view of the shortcomings of existing UAV remote sensing image classification methods for remote sensing image sequences, such as large workload, low efficiency and poor result reliability, so as to efficiently obtain reliable classification image sequences.

[0005] The technical solution adopted by the present invention is:

[0006] A method for quickly generating ground object classification images for UAV remote sensing image sequences, the method comprising the following steps:

[0007] Step S1, obtaining a sequence of high-resolution UAV remote sensing images of the target area, and stitching them together to generate a first mosaic of the target area; for example, the first mosaic is generated by automated mosaic software; wherein the high-resolution UAV remote sensing images may be visible light data, multi-spectral data, etc., and such data is conducive to the classification of land objects;

[0008] Step S2, performing image classification processing on the first mosaic image to obtain an initial object classification map (i.e., a large classification map) of the target area;

[0009] Among them, image classification processing can adopt high-precision classification methods such as supervised classification or machine learning;

[0010] Step S3, carries out adjustment calibration to the image sequence to be classified of the input target area, and records the parameters such as the camera space position (X, Y, Z coordinates) and camera attitude (such as heading inclination, lateral inclination, image rotation angle, etc.) after calibration;

[0011] If the input image sequence to be classified of the target area is consistent with the image sequence of the mosaic, the image sequence to be classified is spliced ​​to generate a second mosaic of the target area;

[0012] If there is a second mosaic, image registration is performed on it and the initial land feature classification map, and the images are resampled to the same image resolution; that is, the second mosaic is geo-registered and spatial resolution is resampled with the initial land feature classification map, and the geographic information of the initial land feature classification map image is fine-tuned with the second mosaic as a reference;

[0013] Step S4, obtaining the object classification map of each frame of the image sequence to be classified:

[0014] Step S401, for the current frame image (i.e., a frame image in the image sequence to be classified), calculate the image scale and the offset of the image center relative to the photographic center, and determine the row and column number of the image center in the initial object classification map;

[0015] Step S402, calculating the plane rotation transformation matrix of the current frame image, and performing plane rotation on the current frame image based on the plane rotation transformation matrix, and recording the corner point coordinates of the minimum circumscribed rectangle of the image after the plane rotation;

[0016] Step S403, extracting the minimum circumscribed rectangular area determined by step S402 from the initial object classification map to generate a sub-classification image; and performing an inverse plane rotation transformation on the sub-classification image based on the plane rotation transformation matrix of the previous frame image, extracting and exporting the central area image, and obtaining the object classification image of the current frame image.

[0017] Based on the above steps, a ground object classification image corresponding one-to-one to the image sequence to be classified is obtained.

[0018] Furthermore, in S401, the image scale and the offsets dx and dy of the image center point relative to the photographic center in the horizontal direction X and the vertical direction Y are calculated according to the focal length, pixel size, calibrated camera posture, and spatial resolution of the second mosaic.

[0019] Furthermore, in S401, according to the recorded horizontal and vertical coordinates of the calibrated camera, and the offsets dx and dy, based on the geographic information of the initial object classification image, the row and column numbers of the center point of the current frame image in the initial object classification image are calculated.

[0020] The technical solution provided by the present invention brings at least the following beneficial effects:

[0021] The present invention can quickly and accurately generate ground object classification images that correspond one-to-one to UAV remote sensing image sequences, overcoming the defects of traditional methods such as large workload, low efficiency, and high uncertainty, thereby providing support for the further application of UAV remote sensing image sequences. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0023] Figure 1 A schematic diagram of a process for rapidly generating a ground object classification image for a UAV remote sensing image sequence provided by an embodiment of the present invention;

[0024] Figure 2 The ground feature classification image of Daman Station on July 13, 2020 provided in an embodiment of the present invention is generated by supervised classification based on the multispectral image of the survey area;

[0025] Figure 3 The brightness temperature orthomosaic map of Daman Station on July 13, 2020 provided in an embodiment of the present invention;

[0026] Figure 4 A diagram of the minimum bounding rectangle area of ​​a frame of an image in a sequence of images to be classified in a brightness temperature mosaic and a classification image provided by an embodiment of the present invention, wherein (a) is a brightness temperature image, (b) is a classification image, and the dotted box in the diagram represents the actual coverage of the image frame on the ground;

[0027] Figure 5A sub-image generated by plane rotation of the minimum circumscribed rectangular area extracted from a frame of an image sequence to be classified provided by an embodiment of the present invention, wherein (a) is a brightness temperature image, (b) is a classification image, and the dotted box in the figure represents the actual coverage of the image frame on the ground;

[0028] Figure 6 The final backtracking generation result for a frame image in the image sequence to be classified provided by an embodiment of the present invention, wherein (a) is the final backtracking generation result of the brightness temperature image, (b) is the final backtracking generation result of the classified image, and (c) is the original thermal infrared image taken by the camera. DETAILED DESCRIPTION

[0029] To make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be described in detail and completely in conjunction with the drawings in the implementation of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the present application for protection, but merely represents selected embodiments of the present invention.

[0030] The embodiment of the present invention discloses a method for quickly generating a ground object classification image for a UAV remote sensing image sequence. First, a large ground object classification map that completely covers the survey area is generated through visible light, multi-spectral and other data that are easy to classify ground objects; then, the UAV remote sensing image sequence to be classified is spliced ​​into an ortho-mosaic map using automated jigsaw software, and the classification large map is aligned and resampled with it, and the calibrated camera spatial position, posture and other parameters are derived at the same time. Next, for a certain frame of the image in the image sequence to be classified, the scale and the offset of the image center relative to the photography center are calculated in combination with the camera's internal orientation elements and the spatial resolution of the ortho-mosaic map, so as to determine its row and column number in the classification large map. Finally, the corresponding classification sub-map can be extracted from the classification large map in combination with the size of the image frame and the film rotation angle; the corresponding ground object classification image sequence can be quickly generated by traversing each frame of the remote sensing image sequence to be classified. The embodiment of the present invention starts from the perspective of "image backtracking", fully considers factors such as the spatial position, posture and rotation changes of the camera, and quickly and accurately obtains the ground object classification image sequence corresponding to the remote sensing image sequence, overcoming the defects of the traditional method such as large workload, low efficiency and low accuracy; in addition, since "image backtracking" is not only applicable to ground object classification data, but also to important parameter data such as surface emissivity and vegetation index; therefore, the method of the present invention has good scalability and practicality, and can serve a large number of UAV remote sensing application scenarios.

[0031] As a possible implementation, Figure 1 As shown, the method for quickly generating ground object classification images for UAV remote sensing image sequences provided by the embodiment of the present invention includes 10 steps: generating an orthophoto mosaic that completely covers the survey area for ground object classification; performing classification based on the mosaic to obtain a large ground object classification image; calibrating the position and attitude of the image sequence to be classified and recording relevant parameters; splicing the image sequence to be classified to generate a new orthophoto mosaic; aligning the large classification image with the new mosaic, and resampling to the same resolution; calculating the image frame scale and the offset of the image center point relative to the photography center, and determining the row and column number of the image center in the large classification image; calculating the plane rotation transformation matrix, and obtaining the coordinates of the corner points of the minimum circumscribed rectangle; extracting the sub-classification image corresponding to the minimum circumscribed rectangle from the large classification image; performing an inverse rotation transformation on the sub-classification image, extracting the central area image and exporting it; traversing each frame image in the image sequence to be classified and repeatedly processing it, so as to generate a ground object classification image sequence corresponding to the original image sequence, and the specific implementation process of each step is as follows:

[0032] Step 1: Generate an orthomosaic map that completely covers the survey area for feature classification.

[0033] A visible light camera or a multispectral camera is used to obtain a remote sensing image sequence that completely covers the survey area. This is because such images have high resolution, rich image details, and a large number of bands, which is conducive to subsequent classification operations. The embodiment of the present invention takes the multispectral data acquired at Daman Station in Zhangye City, Gansu Province on July 13, 2020 as an example to generate a basic image for classification. The data is captured by a RedEdge-M multispectral camera with five bands: red, green, blue, near infrared, and red edge. A single frame image has 1920×1080 pixels and a spatial resolution of approximately 0.2m. Image stitching is performed using Pix4D mapper automated puzzle software.

[0034] Step 2: Obtain a large map of the land feature classification in the survey area.

[0035] Reliable classification basic data can be obtained through step 1. The embodiment of the present invention adopts the supervised classification method to set the ground object categories in the survey area into five categories: cornfield, grassland, bare land, trees, and buildings. Classification is performed using ENVI software; first, a large number of labeled samples should be evenly selected for each category on the mosaic map, and the separability should be calculated using the software. If the separability is too low, the samples need to be reselected; then, an appropriate classification criterion is selected for automatic classification. The embodiment of the present invention adopts the maximum likelihood criterion, as shown in formula (1):

[0036] Ln(L(x))=Ln(p(C i |x))=Ln(p(x|C i )p(C i ))=Ln(p(x|Ci ))+Ln(p(C i )) (1)

[0037] In the formula, p() represents the probability; x represents the sample of the undetermined category; C i represents the i-th classification category; L() represents the objective function; Ln() represents the logarithmic operation.

[0038] Figure 2 The classification results in the embodiment of the present invention are shown. It can be seen that the distribution of ground object categories conforms to the actual situation in the survey area, and there are fewer phenomena such as spots and patches in the figure. According to the calculation results of the confusion matrix, it can be seen that the overall classification accuracy is high, better than 90%.

[0039] Step 3: Calibrate the position and posture of the image sequence to be classified and record the relevant parameters.

[0040] Since the "image backtracking" process of the image sequence to be classified requires matching the position and posture of the image frame from the classification map, it is crucial to obtain the corrected position and posture parameters. The position and posture of the image sequence are usually globally optimized using the bundle adjustment method, and automated puzzle software can usually complete this operation. The embodiment of the present invention uses Pix4Dmapper software to optimize the position and posture data of the image sequence to be classified and export it as a txt file for subsequent operations; the specific content of the file includes: image name, X, Y, Z coordinates, as well as heading inclination, lateral inclination, and film rotation angle data, and synchronously records parameters such as focal length and pixel resolution.

[0041] Step 4: Stitch the image sequence to be classified to generate a new orthomosaic.

[0042] Since the image sequence to be classified is not necessarily the same as the image sequence used to generate the large classification image, there is a mismatch between the two in time and space, and it is difficult to directly perform "image backtracking" matching between the two; therefore, it is necessary to use the orthophoto mosaic generated by the image sequence to be classified as an "intermediate bridge" for subsequent operations. The image sequence to be classified in the embodiment of the present invention is the thermal infrared remote sensing data of a drone, which is taken by a WIRIS Pro Sc thermal infrared camera. A single frame image has 640×512 pixels and a spatial resolution of approximately 0.4m. The shooting area is the same as that of the aforementioned multispectral camera, and the puzzle is automatically processed using the Pix4D mapper software. Figure 3 The stitched brightness temperature orthomosaic is shown.

[0043] Step 5: Align the classified large image with the new mosaic and resample to the same resolution.

[0044] Using ArcMap software, the classification map in step 2 is georeferenced with the brightness temperature orthophoto mosaic in step 4. Considering that the flight path of the UAV is relatively regular and the camera posture usually changes within a small range, a transformation model is used for registration (with the new mosaic, i.e., the brightness temperature orthophoto mosaic, as a reference); this registration method mainly considers the translation, rotation, and scaling changes between images, and the transformation principle can be described by formula (2). In order to ensure that the image sequence to be classified can be traced back to the correct position, it is also necessary to sample the spatial resolution of the classification map to the spatial resolution of the new mosaic; considering that the classification data is a discrete value, the nearest neighbor method is used to sample it, and the principle is shown in formula (3).

[0045]

[0046] Where, T scale 、T rotate 、T trans They represent scaling, rotation, and translation transformation matrices respectively; a and b represent scaling parameters; θ represents the rotation angle, with the counterclockwise direction being positive; t x and t y Respectively represent the translation amount in the x and y directions; x and y represent the coordinates before transformation; u and v represent the coordinates after transformation.

[0047] P(u,v)=Q(x,y) (3)

[0048] Where Q(x,y) represents the pixel value in the original image with the smallest distance to the transformed image coordinate (u,v); P(u,v) represents the pixel value at the transformed image coordinate (u,v).

[0049] After the operation of formulas (2)-(3), a large classification image matching the new mosaic can be obtained, which is convenient for the subsequent "image tracing" process.

[0050] Step 6: Calculate the image frame scale and the offset of the image center relative to the photography center, and determine the row and column number of the image center in the classification map

[0051] During the flight of the drone, the camera is usually not strictly vertically downward, but tilted in the heading and lateral directions, which will cause the center of the captured image to not coincide with the projection of the camera's photographic center on the XOY plane. Therefore, in order to accurately find the position of the image center in the classification image, it is necessary to calculate its displacement relative to the photographic center; the calculation of the displacement is related to the image scale, camera focal length, lateral tilt angle, and heading tilt angle, as shown in formulas (4)-(6):

[0052] scale=mosaic_res / frame_res (4)

[0053] Where mosaic_res represents the spatial resolution of the classification image, in m; frame_res represents the resolution of each pixel of the current image frame, in μm; scale represents the image scale.

[0054]

[0055] Where, f represents the focal length of the camera, in meters; It represents the lateral tilt angle, in degrees, with clockwise direction being positive; dx represents the horizontal displacement of the center point of the current image frame relative to the photography center, in meters.

[0056]

[0057] Where ω represents the heading inclination angle, the unit is degree, the counterclockwise direction is positive, and dy represents the vertical displacement of the center point of the current image frame relative to the photography center.

[0058] The geographic coordinates of the center point of the current image frame in the classification map can be calculated by formulas (4)-(5), and the row and column numbers of the image frame in the classification map can be determined by combining the image resolution.

[0059] Step 7: Calculate the plane rotation transformation matrix and obtain the coordinates of the corner points of the minimum circumscribed rectangle.

[0060] Since step 6 only determines the position (row and column number) of the coordinates of the center point of the current image frame in the classification image, the corresponding actual observation range needs to be further calculated in combination with the camera posture, as shown in formula (7):

[0061]

[0062] Where, κ represents the image rotation angle, the unit is degree, and the counterclockwise direction is positive; col original and row original They represent the original coordinates of the corner points of the image frame in the Cartesian coordinate system with the center of the current image frame as the origin (corresponding to the column number and row number respectively); col rotated and row rotated They respectively represent the coordinates of the image frame corner points after rotation transformation in the Cartesian coordinate system with the center of the current image frame as the origin (corresponding to the column number and row number, respectively).

[0063] It is also necessary to further convert the coordinates obtained by formula (7) into row and column numbers in the classification graph, as shown in formula (8):

[0064]

[0065] In the formula, col mosaic Indicates the column number of the corner point of the rotated image frame in the classification map; rowcenter Indicates the row number of the center point of the rotated image frame in the classification map; row mosaic Indicates the row number of the corner point of the rotated image frame in the classification map.

[0066] Further, the corner point coordinates of the minimum enclosing rectangle can be determined according to the calculation result of the row and column numbers of the four corner points of the original image frame by formula (8); the coordinates of the upper left corner point and the lower right corner point of the rectangle can be expressed by formula (9):

[0067]

[0068] In the formula, row mosaic Represents the vector of row numbers of the corner points of the rotated image frame in the classification map; col mosaic Represents the vector composed of the column numbers of the corner points of the rotated image frame in the classification map; P UL Indicates the coordinates of the upper left corner of the minimum circumscribed rectangle (row and column numbers in the classification map); P LR Indicates the coordinates of the lower right corner of the minimum enclosing rectangle (row and column numbers in the classification image).

[0069] Step 8: Extract the sub-category image corresponding to the minimum bounding rectangle from the classification image.

[0070] The minimum bounding rectangle calculated by formula (9) can be used to crop the corresponding sub-classification image from the classification image. Figure 4 The minimum bounding rectangle extraction result of a frame in the brightness temperature image sequence to be classified by the embodiment of the present invention is shown. Considering that it is difficult to see the details of the classified image, Figure 4 (a) shows the extraction results of the brightness temperature mosaic map for comparison; Figure 4 (b) shows the extraction result for the large classification image. It can be found that the bounding rectangle correctly contains the range of the original image frame, and the position deviation is small.

[0071] Step 9: Perform inverse rotation transformation on the sub-classified image, extract the central area image and export it.

[0072] In order to obtain a classification sub-image with the same orientation as the original image frame, it is also necessary to perform an inverse rotation transformation on the sub-classification image extracted in step 8 around its center. This process can be expressed by formula (10):

[0073]

[0074] Where x0 and y0 represent the horizontal and vertical coordinate sequences before the transformation, respectively; x′ and y′ represent the horizontal and vertical coordinate sequences after the transformation, respectively.

[0075] Figure 5(b) shows the inverse rotation transformation result of the sub-image obtained in step 8 according to the embodiment of the present invention ( Figure 5 (a) brightness temperature image is used as a reference for comparison). It can be seen that the image posture is well restored after the inverse transformation, and the rotation angle of the rectangle in the middle is basically zero. Further combined with the number of pixels in the width and height of the original image, a rectangle of the same size can be extracted from the center of the inversely rotated image as the result of "image tracing", thereby generating a classification thumbnail. Figure 6 (b) shows the classification thumbnail finally generated by the embodiment of the present invention ( Figure 6 (a) brightness temperature image as a reference for comparison), it can be seen that Figure 6 (c) The original brightness temperature image has a high degree of consistency, and the classification image generation effect is good.

[0076] Step 10: Generate a ground object classification image sequence corresponding to the original remote sensing image sequence.

[0077] Repeat steps 6-9 until all images to be classified are traversed, and a ground object classification image sequence corresponding to the original remote sensing image sequence can be generated.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0079] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the creative concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A method for quickly generating ground object classification images for UAV remote sensing image sequences, characterized in that: The following steps are involved: Step S1, obtaining a high-resolution UAV remote sensing image sequence of the target area, and splicing the image sequence to generate a first mosaic of the target area; Step S2, performing image classification processing on the first mosaic image to obtain an initial ground object classification map of the target area; Step S3, performing adjustment calibration on the image sequence to be classified in the input target area, and recording the calibrated camera space position and camera posture; If the input image sequence to be classified of the target area is inconsistent with the image sequence of the mosaic, the image sequence to be classified is spliced ​​to generate a second mosaic of the target area; If there is a second mosaic, it is image registered with the initial feature classification map and resampled to the same image resolution; Step S4, obtaining the object classification map of each image frame of the image sequence to be classified: Step S401, for the current image frame, calculate the image scale and the offset of the image center relative to the photographic center, and determine the row and column numbers of the image center in the initial object classification map; Step S402, calculating the plane rotation transformation matrix of the current image frame, performing plane rotation on the current image frame based on the plane rotation transformation matrix, and recording the corner coordinates of the minimum circumscribed rectangle of the image frame after the plane rotation; Step S403, extracting the minimum circumscribed rectangular area determined in step S402 from the initial object classification map to generate a sub-classification image; And based on the plane rotation transformation matrix of the current image frame, the sub-classification image is subjected to an inverse plane rotation transformation, the central area image is extracted and exported, and the ground object classification image of the current image frame is obtained.

2. The method according to claim 1, characterized in that The image sequence is puzzled by automated puzzle software to obtain the corresponding mosaic image.

3. The method according to claim 1, characterized in that In step 1, the high-resolution UAV remote sensing image used is visible light data or multispectral data.

4. The method according to claim 1, characterized in that In step S2, the image classification process may adopt a supervised classification method or a machine learning classification method.

5. The method according to claim 1, characterized in that In step S3, the camera posture includes: heading tilt angle, lateral tilt angle and image rotation angle.

6. The method according to claim 5, characterized in that In step S401, the image scale and the offsets dx and dy of the image center point relative to the photographic center in the horizontal direction X and the vertical direction Y are calculated according to the focal length, pixel size, calibrated camera posture and spatial resolution of the second mosaic.

7. The method according to claim 5, characterized in that In S401, according to the recorded horizontal and vertical coordinates of the calibrated camera, and the offsets dx and dy of the image center point relative to the photography center in the horizontal direction X and the vertical direction Y, based on the geographic information of the initial object classification image, the row and column numbers of the center point of the current image frame in the initial object classification image are calculated.

8. The method according to claim 5, characterized in that In step S402, the plane rotation transformation matrix is: Among them, κ represents the rotation angle of the film, and the counterclockwise direction is positive, col original and row original The column and row numbers represent the original coordinates of the corner point of the current image frame in the Cartesian coordinate system with the center of the current image frame as the origin; col rotated and row rotated They respectively represent the column number and row number of the coordinates of the corner point of the current image frame after rotation transformation in the Cartesian coordinate system with the center of the current image frame as the origin.

9. The method according to claim 5, characterized in that In step S402, the corner coordinates of the minimum circumscribed rectangle of the image frame after plane rotation are: Among them, row mosaic Represents the vector consisting of the row numbers of the corner points of the rotated image frame in the initial object classification map, col mosaic Represents the vector composed of the column numbers of the corner points of the rotated image frame in the initial object classification map, P UL P represents the coordinates of the upper left corner of the minimum circumscribed rectangle in the initial feature classification map; LR Represents the coordinates of the lower right corner of the minimum circumscribed rectangle in the initial feature classification map.

Images