A method for obtaining high-definition images by approaching a drone

Abstract

The present invention discloses a method for obtaining high-definition images by approaching a drone, which relates to the fields of drone technology and high-definition image acquisition technology, and comprises the following steps: Step S1: collecting high-resolution image data of the target area by means of a multispectral imaging device carried by the drone, and synchronously recording the flight attitude parameters of the drone. The present invention replaces the reliance on preset terrain data through multispectral imaging and dynamic depth of field analysis, enabling the drone to autonomously plan paths in unknown terrains, thereby improving environmental adaptability. Secondly, dynamic path planning combined with real-time parameter adjustment can cope with dynamic obstacles in complex environments, avoid manual reshooting, and improve work efficiency. Furthermore, multi-scale feature extraction and fusion technology, combined with an improved geometric correction algorithm, can generate high-precision panoramic images, eliminate distortion, and ensure image quality. In addition, strategies such as adaptive PID control ensure the smooth flight of the drone, further improving image clarity and mission success rate.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle (UAV) technology and high-definition image acquisition technology, and in particular to a method for obtaining high-definition images by approaching a UAV. Background Art

[0002] In many fields today, such as surveying and mapping, geological disaster monitoring, cultural relics protection, forestry surveys, etc., drone technology for obtaining high-definition images at close range has been widely used.

[0003] A search revealed a close-up photogrammetry method for a rotary-wing drone, published on April 10, 2020, with publication number CN110006407B. This patent proposes a close-up photogrammetry method that generates a three-dimensional track based on initial terrain information. The method then acquires ultra-high-resolution images by performing automated flight close to the ground or surface (5 to 30 meters) and supplemented by manual control and handheld supplementary photography.

[0004] However, this technical solution has the following shortcomings in practical applications: First, it relies on initial terrain information reconstructed from known or conventional photographic images. If the initial terrain data is inaccurate or missing, it will have a certain impact on the accuracy of trajectory planning; second, although manual control and handheld re-shooting can make up for the image acquisition needs of areas that the drone cannot reach, it increases the cost and time of manual intervention and reduces overall operational efficiency; in addition, the adaptability of this method in complex environments is relatively limited, especially in scenes with many dynamic obstacles, and the ability to respond quickly and adjust in real time needs to be improved.

[0005] In view of the above problems, the existing technology is in urgent need of improvement. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention provides a method for obtaining high-definition images by approaching with a drone, which solves the problem that the existing technology relies on initial terrain information reconstructed from known or conventional photographic images. If the initial terrain data is inaccurate or missing, it will have a certain impact on the accuracy of trajectory planning; secondly, although manual control and handheld supplementary shooting can make up for the image acquisition needs of areas that the drone cannot reach, it increases the cost and time of manual intervention and reduces the overall operation efficiency; in addition, the adaptability of this method in complex environments is relatively limited, especially in scenes with many dynamic obstacles, and the ability to respond quickly and adjust in real time needs to be improved.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for obtaining high-definition images by approaching with a drone, comprising the following steps: step S1: collecting high-resolution image data of a target area through a multispectral imaging device carried by the drone, and synchronously recording the flight attitude parameters of the drone; performing dynamic depth of field analysis on the collected image data to generate a depth of field distribution map of the target area; step S2: determining the optimal shooting distance of each sub-area in the target area based on the depth of field distribution map, and planning a dynamic flight path that meets the optimal shooting distance in combination with the real-time position information of the drone; step S3: controlling the drone to fly automatically according to the dynamic flight path, and adjusting the camera's focal length and exposure parameters in real time during the flight to ensure that the image clarity of each sub-area in the target area reaches a preset threshold; step S4: performing multi-scale feature extraction and fusion on the collected high-resolution image data to generate an ultra-high-resolution panoramic image of the target area, and performing geometric correction on the panoramic image through a feature matching algorithm to eliminate image distortion.

[0008] Furthermore, step S1 includes the following steps: step S11: deploying a multispectral imaging device on the UAV to collect high-resolution image data of the target area; step S12: synchronously recording the attitude parameters of the UAV during flight, including pitch angle, roll angle and yaw angle, through the built-in inertial measurement unit of the UAV; step S13: performing pixel-level brightness equalization processing on the collected image data to enhance the contrast of the image data and reduce the impact of uneven illumination to obtain brightness-balanced image data; step S14: performing dynamic depth of field analysis on the brightness-balanced image data to calculate the depth of field value corresponding to each pixel point in the target area, and generate a depth of field distribution map of the target area; step S15: inputting the depth of field distribution map into the improved deep learning model to extract significant feature points and their descriptors in the target area, and performing non-maximum suppression on the significant feature points and their descriptors to generate a depth of field feature point set of the target area.

[0009] Furthermore, non-maximum suppression of salient feature points and their descriptors includes: performing local texture feature analysis on the salient feature points and their descriptors, analyzing the neighborhood structure corresponding to the salient feature points in the image, and calculating the gradient amplitude histogram of pixels in the neighborhood of the salient feature points to characterize the edge direction distribution characteristics of the salient feature points, while for the descriptors, analyzing their numerical distribution laws and dimensional correlations in different bands to obtain local characteristic data sets of salient feature points and descriptors; performing preliminary screening of the salient feature points and their descriptors based on the local characteristic data sets of the salient feature points and descriptors, so as to associate the salient feature points and their descriptors in the time dimension considering time continuity. , from which the position changes corresponding to the salient feature points of adjacent frames and the similarities corresponding to the descriptors are analyzed, and the salient feature points with abnormal position changes or the descriptors with descriptor similarity lower than a preset threshold are marked as objects to be excluded for screening and exclusion, thereby obtaining a set of salient feature points and descriptors after preliminary screening; a multi-scale local extreme value region is defined for each point in the set of salient feature points and descriptors after preliminary screening, so as to determine the corresponding local extreme value region according to the pixel texture condition in its neighborhood, thereby obtaining a multi-scale local extreme value region; based on the multi-scale local extreme value region, non-maximum suppression is performed on the set of salient feature points and descriptors after preliminary screening, so as to generate a depth of field feature point set of the target area.

[0010] Furthermore, the multi-scale local extreme value region specifically sets a 3x3 scale local extreme value region in the high texture complexity region to capture detail features, and sets a 7x7 scale local extreme value region in the low texture complexity region to consider the overall features.

[0011] Furthermore, step S2 includes the following steps: step S21: sub-dividing the depth of field feature point set of the target area to generate multiple sub-areas of the target area; step S22: determining the optimal shooting distance range of each sub-area based on the depth of field distribution map, and performing dynamic constraint analysis in combination with the current flight altitude and speed of the UAV to generate a dynamic flight path that meets the optimal shooting distance; step S23: optimizing the dynamic flight path in real time to avoid obstacles that may exist in the path and ensure that the UAV can smoothly transition to the next shooting point; step S24: generating flight control instructions for the UAV according to the optimized dynamic flight path, and adjusting the flight attitude of the UAV in real time to ensure that the camera is always facing the target area.

[0012] Furthermore, step S24 includes the following steps: step S241: decomposing the dynamic flight path into multiple path segments, and assigning corresponding flight attitude parameters to each path segment; step S242: decomposing the flight attitude parameters in each path segment into directional components to obtain the attitude adjustment components of the UAV in the horizontal, vertical and yaw directions; step S243: inversely solving the attitude angle of the UAV based on the attitude adjustment components of the UAV in the horizontal, vertical and yaw directions to obtain the attitude adjustment angle of the UAV between adjacent path segments; step S244: generating flight control instructions for the UAV based on the attitude adjustment angle of the UAV between adjacent path segments and combined with the dynamic flight path to ensure that the UAV can smoothly complete the path switching.

[0013] Furthermore, step S244 includes the following steps: performing displacement change analysis on each path segment in the dynamic flight path in the time dimension to obtain the displacement change value of the UAV between adjacent path segments over time; calculating the displacement time derivative based on the displacement change value of the UAV between adjacent path segments over time to obtain the displacement speed and acceleration of the UAV between adjacent path segments; and controlling the flight of the UAV based on the attitude adjustment angle, displacement speed and acceleration of the UAV between adjacent path segments and in combination with an adaptive PID control algorithm to generate a final flight control instruction.

[0014] Furthermore, step S3 includes the following steps: step S31: matching the real-time position information of the UAV with the dynamic flight path, and calculating the deviation between the current position of the UAV and the target shooting point; step S32: adjusting the flight speed and direction of the UAV in real time based on the deviation information to ensure that the UAV can accurately reach the target shooting point; step S33: after the UAV arrives at the target shooting point, adjusting the focal length and exposure parameters of the camera in real time according to the depth of field distribution map of the target area to ensure that the image clarity of each sub-area in the target area reaches a preset threshold; step S34: verifying the adjusted camera parameters. If the image clarity does not reach the preset threshold, readjusting the camera parameters until the requirements are met.

[0015] Furthermore, step S4 includes the following steps: step S41: performing multi-scale feature extraction on the collected high-resolution image data to generate a multi-scale feature map of the target area; step S42: performing feature fusion on the image data of the target area based on the multi-scale feature map to generate an ultra-high-resolution panoramic image of the target area; step S43: performing geometric correction on the panoramic image through a feature matching algorithm to eliminate image distortion; step S44: generating a final high-definition image of the target area based on the geometrically corrected panoramic image.

[0016] Furthermore, the feature matching algorithm adopts an improved algorithm based on RANSAC, and its formula is as follows: ,in, Indicates the accumulation operation of all N matching point pairs. is the index of the matching point pair (from 1 to N), N is the total number of matching point pairs, Represents the homography matrix H for the Source image feature points The geometric transformation result, H represents the homography matrix, and Represent the coordinates of the matching point pairs, represents the Euclidean distance function, Represents a robust error function, which is used to reduce the influence of mismatched points. The algorithm achieves accurate geometric correction of panoramic images by iteratively optimizing the homography matrix H. Beneficial effects

[0017] This invention replaces the reliance on preset terrain data through multispectral imaging and dynamic depth-of-field analysis, enabling drones to autonomously plan paths in unknown terrain, improving environmental adaptability. Secondly, dynamic path planning combined with real-time parameter adjustment can cope with dynamic obstacles in complex environments, avoid manual reshoots, and improve operational efficiency. Furthermore, multi-scale feature extraction and fusion technology, combined with an improved geometric correction algorithm, can generate high-precision panoramic images, eliminate distortion, and ensure image quality. Furthermore, strategies such as adaptive PID control ensure stable drone flight, further improving image clarity and mission success rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0019] The technical solutions in this application will be clearly and completely described below in conjunction with the drawings in this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all of the embodiments. The components of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for which protection is claimed, but merely represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of this application.

[0020] It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and should not be understood as indicating or implying relative importance.

[0021] In existing technologies, drone close-up photogrammetry often relies on preset terrain information to generate flight paths. When the initial terrain data is missing or the error is large, the accuracy of trajectory planning decreases significantly. Existing methods require manual re-shooting to compensate for image acquisition defects in areas that the drone cannot reach, resulting in reduced operational efficiency. In addition, dynamic obstacles in complex environments are difficult to avoid in real time, affecting image acquisition quality and mission success rate. For example, in landslide monitoring scenarios, areas with sudden changes in terrain often cause drones to deviate from the optimal shooting distance due to a lack of accurate initial data. Manual intervention and re-shooting are time-consuming and labor-intensive, and it is difficult to cope with dynamic obstacles such as real-time changes in debris landslides.

[0022] To address the above issues, the inventors discovered that the key bottlenecks of existing technologies lie in their dependence on static terrain and human intervention. By analyzing the dynamic correlation between image acquisition quality and flight paths, they proposed using depth of field distribution as the core basis for autonomous path planning. Further consideration revealed that multispectral imaging equipment can simultaneously acquire multi-dimensional image data, which, combined with flight attitude parameters, can achieve three-dimensional spatial reconstruction, thereby replacing preset terrain data. In order to adapt to complex environments, the inventors realized that real-time depth of field analysis can dynamically capture terrain changes, while multi-scale feature fusion can eliminate image distortion caused by dynamic flight. This has formed a technical route for achieving closed-loop control through dynamic depth of field perception, autonomous path planning, and real-time parameter adjustment.

[0023] Therefore, this application proposes a method for obtaining high-definition images by approaching a drone, comprising the following steps:

[0024] Step S1: Collect high-resolution image data of the target area through the multispectral imaging equipment carried by the UAV, and simultaneously record the flight attitude parameters of the UAV; perform dynamic depth of field analysis on the collected image data to generate a depth of field distribution map of the target area;

[0025] Step S2: Determine the optimal shooting distance for each sub-area within the target area based on the depth of field distribution map, and plan a dynamic flight path that meets the optimal shooting distance in combination with the real-time position information of the UAV;

[0026] Step S3: Controlling the UAV to fly automatically according to the dynamic flight path, and adjusting the focus and exposure parameters of the camera in real time during the flight to ensure that the image clarity of each sub-area within the target area reaches a preset threshold;

[0027] Step S4: Multi-scale feature extraction and fusion are performed on the collected high-resolution image data to generate an ultra-high-resolution panoramic image of the target area, and the panoramic image is geometrically corrected through a feature matching algorithm to eliminate image distortion.

[0028] Specifically, high-resolution image data of the target area is collected through the multispectral imaging equipment carried by the drone, and the flight attitude parameters are recorded simultaneously; dynamic depth of field analysis is performed on the image data to generate a depth of field distribution map; based on the depth of field distribution map, the optimal shooting distance for each sub-area is determined and a dynamic flight path is planned; the drone is controlled to fly along the path and the camera parameters are adjusted in real time; and a technical solution is developed for multi-scale feature extraction, fusion and geometric correction of the image data to generate a panoramic image.

[0029] Among them, multispectral imaging equipment refers to a camera device with multi-band imaging capabilities such as visible light and near-infrared. It can be implemented by a modular device that integrates multiple spectral sensors to simultaneously obtain high-resolution images of different spectral characteristics of the target area. Dynamic depth of field analysis refers to an algorithm that calculates the depth of field value based on the relationship between the brightness and spatial position of image pixels. It can be implemented by binocular parallax matching combined with a deep learning model to construct a three-dimensional structural model of the target area. The optimal shooting distance range refers to the distance interval between the drone and the object to be photographed that ensures image clarity. It can be determined by calculating the geometric relationship between the depth of field value and the focal length of the camera, and is used to guide path planning. Multi-scale feature extraction and fusion refers to the technology of extracting and integrating image features at different resolution levels. It can be implemented by a convolutional neural network combined with a pyramid structure to improve image resolution and detail restoration capabilities.

[0030] Specifically, the drone continuously collects multispectral image data during flight, simultaneously recording attitude parameters such as pitch and roll angles. A dynamic depth-of-field analysis algorithm processes the raw images to generate a depth distribution map, quantifying the three-dimensional characteristics of the target area. The path planning module divides the drone into sub-areas based on the depth distribution, calculates the optimal shooting distance for each area, and generates an obstacle avoidance path based on the real-time position. The flight control system adjusts the drone's attitude based on the path instructions, synchronously linking the camera's focal length and exposure parameters to ensure image quality at different depths of field. Multi-scale feature extraction is performed on the captured images to integrate spatial details, and geometric correction is then used to eliminate distortion caused by flight attitude changes, ultimately producing a high-precision panoramic image. This enables the drone to autonomously complete close-up photography missions in unknown terrain environments, eliminating reliance on pre-set terrain data. Dynamic path planning and real-time parameter adjustment enable the drone to adapt to complex environmental changes and avoid manual re-shooting. Multispectral data and multi-scale processing technologies ensure that image acquisition quality meets surveying and mapping requirements, and the geometric correction algorithm effectively eliminates image distortion caused by flight attitude changes. The entire system forms a complete closed loop from data acquisition to image output, improving operational efficiency and reliability.

[0031] This application further proposes a technical solution comprising the following steps:

[0032] Step S11: deploying a multispectral imaging device on the UAV to collect high-resolution image data of the target area;

[0033] Step S12: Synchronously record the attitude parameters of the drone during flight, including pitch angle, roll angle, and yaw angle, through the drone's built-in inertial measurement unit;

[0034] Step S13: performing pixel-level brightness equalization processing on the collected image data to enhance the contrast of the image data and reduce the influence of uneven illumination, thereby obtaining brightness-balanced image data;

[0035] Step S14: performing dynamic depth of field analysis on the brightness balanced image data to calculate the depth of field value corresponding to each pixel point in the target area and generate a depth of field distribution map of the target area;

[0036] Step S15: Input the depth distribution map into the improved deep learning model to extract the significant feature points and their descriptors in the target area, and perform non-maximum suppression on the significant feature points and their descriptors to generate a depth feature point set of the target area.

[0037] Multispectral imaging equipment refers to sensors capable of simultaneously collecting image data from multiple bands. This can be achieved using multispectral cameras or spectral spectrometers, enhancing the richness of image information through multi-band data fusion. An inertial measurement unit (IMU) is a device used to measure the three-dimensional motion parameters of a drone. It can be implemented using a sensor module combining a gyroscope and accelerometer. Real-time recording of flight attitude parameters ensures accurate correspondence between image data and spatial position. Pixel-level brightness equalization is an algorithm that normalizes the brightness of each pixel in an image. This can be achieved using histogram equalization or adaptive gamma correction. This algorithm improves overall image contrast by eliminating overexposed or underexposed areas. Dynamic depth of field analysis is the process of calculating scene depth information based on image data. This can be achieved using binocular disparity matching or deep learning model prediction methods. This method generates a three-dimensional spatial distribution map of the target area through pixel-by-pixel depth estimation. Non-maximum suppression is a method for filtering out redundant feature points within a local area. This can be achieved using gradient magnitude comparison or feature response value sorting algorithms. This algorithm reduces the complexity of subsequent processing by retaining the most significant feature points.

[0038] Specifically, high-resolution image data collected by multispectral imaging equipment overcomes the illumination sensitivity issues associated with single-spectrum imaging through the complementary use of multi-band information. The pitch, roll, and yaw angles recorded by the inertial measurement unit are synchronized with the image acquisition timestamp, providing spatial attitude compensation parameters for subsequent depth-of-field analysis. Pixel-level brightness equalization addresses the contrast loss caused by illumination differences in images by adaptively adjusting the pixel brightness distribution, effectively enhancing both dark details and highlights. Dynamic depth-of-field analysis uses the equalized image data and drone attitude parameters for three-dimensional spatial mapping, generating an accurate depth distribution map reflecting the three-dimensional structural characteristics of the target area. An improved deep learning model extracts spatially distinct and salient feature points by fusing multi-scale features with depth-of-field information. A non-maximum suppression algorithm eliminates duplicate or low-distinctive feature points through local extreme value filtering, thereby generating a highly robust depth-of-field feature point set.

[0039] Through the above technical solution, this application solves the problem of insufficient image contrast caused by uneven lighting when the drone approaches for shooting, and improves the image quality through multi-spectral data fusion and brightness equalization processing; realizes autonomous depth of field analysis without relying on external terrain data, and generates accurate three-dimensional distribution maps through dynamic depth calculation; uses a non-maximum suppression algorithm to screen highly significant feature points, reducing the interference of redundant features on path planning, thereby providing a high-precision spatial reference basis for the generation of dynamic flight paths of drones.

[0040] This application further proposes a method for non-maximum suppression of salient feature points and their descriptors, comprising the following steps:

[0041] The local texture characteristics of salient feature points and their descriptors are analyzed to analyze the neighborhood structure corresponding to the salient feature points in the image. The gradient amplitude histogram of the pixels in the neighborhood of the salient feature points is calculated to characterize the edge direction distribution characteristics of the salient feature points. For the descriptors, their numerical distribution laws and dimensional correlations in different bands are analyzed to obtain the local characteristic data set of salient feature points and descriptors.

[0042] Based on the local characteristic dataset of salient feature points and descriptors, the salient feature points and their descriptors are preliminarily screened. The salient feature points and their descriptors are associated in the time dimension considering temporal continuity. The position changes of the salient feature points in adjacent frames and the similarity of the descriptors are analyzed. The salient feature points with abnormal position changes or the descriptors with descriptor similarity lower than a preset threshold are marked as objects to be excluded for screening and exclusion, thereby obtaining a set of salient feature points and descriptors after the preliminary screening.

[0043] A multi-scale local extreme value region is defined for each point in the salient feature points and descriptor set after the initial screening, so as to determine the corresponding local extreme value region according to the pixel texture condition in its neighborhood, and obtain the multi-scale local extreme value region;

[0044] Based on the multi-scale local extreme value region, non-maximum suppression is performed on the salient feature points and descriptor sets after the initial screening to generate the depth of field feature point set of the target area.

[0045] Local texture feature analysis involves extracting the edge direction characteristics of feature points and the band correlation of descriptors through gradient amplitude histogram and multi-band numerical distribution analysis. This can be achieved by using the Sobel operator to calculate pixel gradient amplitude and performing multi-band numerical statistics in the HSV color space. This analysis enhances the distinction between feature points and descriptors, providing multidimensional data support for subsequent screening. Multi-scale local extreme value regions refer to detection windows that are dynamically adjusted based on the complexity of neighborhood textures. This can be achieved by using an adaptive threshold segmentation method to divide high- and low-texture regions, setting 3x3 and 7x7 detection windows, respectively. This definition balances the detail preservation and noise resistance of feature detection, avoiding feature omissions or false detections caused by a single scale.

[0046] Specifically, this method first establishes the edge direction fingerprint of the feature point through the gradient amplitude histogram, and at the same time analyzes the numerical distribution pattern of the descriptor in different spectral bands to form a feature data set containing spatiotemporal correlation information. Subsequently, based on the position change vector and descriptor similarity matrix between adjacent frames, a feature point motion trajectory model is established, and feature points that deviate from the trajectory beyond the set tolerance range or have insufficient similarity are removed. Further, the detection window scale is dynamically selected according to the texture complexity. A small window is used to capture detailed features in high-frequency texture areas, and a large window is used to maintain feature continuity in low-frequency areas. Finally, through the multi-window non-maximum suppression operation, the feature points with the largest response values ​​at each scale are retained, forming a feature point set with uniform spatial distribution and strong geometric consistency. Therefore, this scheme can autonomously identify and eliminate abnormal feature points caused by lighting changes or object movement by real-time analysis of the motion patterns of feature points in adjacent frames, significantly improving the robustness of feature matching in complex environments.

[0047] Through the above technical solution, this application solves the problem of feature point mismatching caused by dynamic changes in the environment when the drone approaches for shooting. Abnormal feature points are eliminated by constraining spatiotemporal continuity, and the feature space distribution is optimized by combining multi-scale extreme value detection, so that the generated depth of field feature point set can still maintain high matching accuracy in texture-repeated areas. In complex scenes such as shooting building facades, this technical solution can effectively avoid feature point drift caused by reflections from glass curtain walls or occlusions from vegetation, providing a reliable feature basis for subsequent image stitching and three-dimensional reconstruction.

[0048] This application further proposes a multi-scale local extreme value region, specifically setting a 3x3 scale local extreme value region in the high texture complexity region to capture detail features, and setting a 7x7 scale local extreme value region in the low texture complexity region to consider the overall features.

[0049] Among them, the high texture complexity area refers to the area in the image where the pixel brightness changes frequently and the gradient amplitude is densely distributed. Specifically, it can be achieved by the gradient amplitude variance in the pixel neighborhood exceeding the preset threshold. This area needs to be analyzed through a small-scale neighborhood to avoid noise interference. The low texture complexity area refers to the area in the image where the pixel brightness changes smoothly and the gradient amplitude is sparsely distributed. Specifically, it can be achieved by the gradient amplitude variance in the pixel neighborhood being lower than the preset threshold. This area needs to be analyzed through a large-scale neighborhood to avoid feature sparseness. The 3x3 scale local extreme value area refers to the neighborhood range composed of three rows and three columns of pixels centered on the target pixel. Specifically, it can be achieved by traversing the image data with a sliding window, and is used to extract local detail features in complex textures. The 7x7 scale local extreme value area refers to the neighborhood range composed of seven rows and seven columns of pixels centered on the target pixel. Specifically, it can be achieved by traversing the image data with an extended sliding window, and is used to extract the overall characteristics of the region in flat textures.

[0050] Specifically, by setting local extreme value areas of different scales and using a 3x3 neighborhood range for areas with high texture complexity, it is possible to focus on subtle brightness differences between pixels, avoid over-crowding of feature points or noise interference due to complex textures, and accurately extract detailed features such as edges and corners. For areas with low texture complexity, a 7x7 neighborhood range is used, which can cover a larger range of pixel brightness distribution. By analyzing the overall change trend of pixels in the neighborhood, it avoids sparse feature points due to single texture, and thus extracts regional contours or structural features. This scheme achieves a balance between detail preservation and overall structural analysis during the feature extraction process by dynamically matching texture complexity with neighborhood scale. Therefore, this scheme solves the problem of insufficient adaptability of fixed-scale neighborhoods in complex scenes by adaptively adjusting the neighborhood scale according to texture complexity, and significantly improves the accuracy and robustness of feature extraction.

[0051] Through the above technical solution, this application effectively solves the problem of difficulty in balancing detail features and overall features in different texture areas. By dynamically adjusting the scale of local extreme areas, it not only ensures the detail capture accuracy of high-texture areas, but also enhances the feature integrity of low-texture areas, thereby providing a more reliable feature data foundation for subsequent image fusion and geometric correction.

[0052] This application further proposes a technical solution comprising the following steps:

[0053] Step S21: Dividing the depth feature point set of the target area into sub-areas to generate multiple sub-areas of the target area;

[0054] Step S22: determining the optimal shooting distance range for each sub-area based on the depth of field distribution map, and performing dynamic constraint analysis in combination with the current flight altitude and speed of the UAV to generate a dynamic flight path that meets the optimal shooting distance;

[0055] Step S23: Optimize the dynamic flight path in real time to avoid obstacles that may exist in the path and ensure that the drone can smoothly transition to the next shooting point;

[0056] Step S24: Generate flight control instructions for the UAV according to the optimized dynamic flight path, and adjust the flight attitude of the UAV in real time to ensure that the camera is always facing the target area.

[0057] Subregion segmentation refers to spatially segmenting the target area based on the distribution of depth-of-field feature points. This can be achieved using density-based clustering or grid-based segmentation algorithms. For example, the DBSCAN algorithm groups feature points with similar depths of field into the same subregion, thereby resolving the issue of the global path being unable to adapt to local depth-of-field differences. The optimal shooting distance range refers to calculating the minimum and maximum shooting distances required to maintain image clarity in each subregion based on the depth-of-field distribution map. This can be achieved using a mathematical model based on the relationship between depth-of-field values ​​and focal length. For example, a depth-of-field formula combined with camera parameters can be used to calculate the focusable range of each subregion, thereby ensuring image acquisition quality in each subregion. Dynamic constraint analysis combines the drone's real-time flight state parameters with the optimal shooting distance to generate a feasible path. This can be achieved using a multi-objective optimization algorithm, such as incorporating flight altitude and speed limits as constraints into the path planning model to balance image capture requirements with the vehicle's maneuverability. Real-time optimization refers to dynamically adjusting the flight path based on environmental changes. This can be achieved using obstacle detection and replanning algorithms. For example, after identifying obstacles using a lidar or visual sensor, the A* algorithm can be used to generate a detour path, thereby improving path adaptability in complex environments. Among them, flight control instruction generation refers to converting the path into attitude adjustment instructions that can be executed by the drone. This can be achieved using an inverse kinematics algorithm. For example, by decomposing the displacement and angle changes of the path segments, the corresponding motor speed and servo deflection are calculated, thereby achieving precise control of the flight attitude.

[0058] Specifically, by dividing the target area into multiple sub-areas, the optimal shooting distance can be calculated independently for each sub-area, avoiding local image blur caused by a single shooting parameter. Dynamic constraint analysis is performed based on the depth of field distribution map and the real-time flight parameters of the drone to generate a path that takes into account both shooting quality and flight stability. During the path execution process, obstacles are detected in real time by sensors and path replanning is triggered to eliminate the risk of collision. At the same time, by decomposing the optimized path into attitude adjustment components and combining the inverse kinematic solution to generate control instructions, the camera is ensured to always be aimed at the target area during flight, reducing image distortion caused by attitude offset. Therefore, this solution achieves a smooth transition of the flight process by linking the decomposition of attitude adjustment components with control instructions.

[0059] Through the above technical solution, this application solves the problem of coordinating drone path planning and image acquisition in complex scenarios. It can automatically generate and optimize flight paths without human intervention, ensuring that image clarity in each sub-area meets the required standards. At the same time, through real-time obstacle avoidance and stable attitude control, the risk of flight mission failure is reduced, and the reliability of operations in complex environments and the efficiency of image acquisition are improved.

[0060] This application further proposes:

[0061] Step S24 includes the following steps:

[0062] Step S241: decomposing the dynamic flight path into multiple path segments, and assigning corresponding flight attitude parameters to each path segment;

[0063] Step S242: Decomposing the flight attitude parameters in each path segment into directional components to obtain attitude adjustment components of the UAV in the horizontal, vertical, and yaw directions;

[0064] Step S243: performing an inverse solution on the attitude angle of the UAV based on the attitude adjustment components of the UAV in the horizontal, vertical, and yaw directions to obtain the attitude adjustment angle of the UAV between adjacent path segments;

[0065] Step S244: Generate flight control instructions for the drone based on the attitude adjustment angle between adjacent path segments and in combination with the dynamic flight path to ensure that the drone can smoothly complete the path switching.

[0066] Path segmentation involves dividing a continuous dynamic flight path into multiple logically independent subsegments. This can be achieved using an adaptive segmentation algorithm based on trajectory curvature changes. This segmentation is performed by detecting extreme points of path curvature or by using a preset segment length threshold, allowing each subsegment to be independently assigned flight attitude parameters. Directional component decomposition involves decomposing flight attitude parameters into horizontal motion components, vertical altitude components, and heading deflection components in a three-dimensional coordinate system. This can be achieved using orthogonal projection or vector decomposition methods. Separating motion parameters from different dimensions enables independent control of multiple degrees of freedom. Attitude angle inversion involves calculating the attitude angle of the drone's body coordinate system relative to the target path segment based on the decomposed directional components. This can be achieved using quaternion conversion or inverse Euler angle calculations. The attitude adjustment between adjacent path segments is calculated by establishing an inverse kinematic model. Flight control command generation involves integrating the attitude adjustment angle with the path parameters to generate executable heading, altitude, and speed commands for the drone. This can be achieved using a PID control algorithm or model predictive control. Dynamic adjustments during path switching are achieved through a closed-loop feedback mechanism.

[0067] Specifically, after the dynamic flight path is decomposed into multiple path segments, the flight attitude parameters of each segment are independently assigned, allowing the UAV's motion characteristics in different path segments to be controlled in a targeted manner. By decomposing the attitude parameters into components in the horizontal, vertical, and yaw directions, the flight control system can separately handle motion adjustment requirements in different dimensions, avoiding the multi-degree-of-freedom coupling interference caused by single parameter adjustments. When performing attitude angle inversion based on the decomposed components, the attitude angle differences between adjacent path segments are reversely deduced by establishing a UAV kinematic model, providing precise adjustment input for path switching. The resulting flight control instructions dynamically combine the attitude adjustment angle with the path parameters, achieving a smooth transition for path switching by adjusting the UAV's roll, pitch, and yaw movements in real time, thereby eliminating the flight jitter or trajectory deviation problems caused by overall attitude mutations in traditional methods. Therefore, this solution significantly improves the accuracy and response speed of attitude adjustment by combining orthogonal decomposition with kinematic inversion.

[0068] Through the above technical solution, this application solves the problem of reduced flight stability of drones due to unstable attitude adjustment during dynamic path switching, and realizes smooth transition control between path segments. Through multi-dimensional parameter decomposition and inverse calculation, the precise matching of attitude adjustment angles and path parameters is ensured, avoiding the flight jitter phenomenon caused by overall parameter adjustment in traditional methods. After adopting the segmented control strategy, the flight stability of drones under complex paths is effectively improved, while the frequency of manual intervention is reduced and the efficiency of automated operations is improved.

[0069] This application further proposes:

[0070] Perform displacement change analysis on each path segment in the dynamic flight path in the time dimension to obtain the displacement change value of the UAV between adjacent path segments over time;

[0071] The displacement time derivative is calculated based on the change in the displacement of the UAV between adjacent path segments over time to obtain the displacement velocity and acceleration of the UAV between adjacent path segments.

[0072] The flight control of the UAV is performed based on the attitude adjustment angle, displacement velocity and acceleration between adjacent path segments and combined with an adaptive PID control algorithm to generate the final flight control instructions.

[0073] Analysis of displacement changes in the temporal dimension refers to quantifying the displacement changes of the drone between path segments using time series data. This can be achieved using discrete-time sampling methods, for example, by collecting the drone's three-dimensional coordinate data at 0.1-second intervals. This establishes a functional relationship between displacement and time, reflecting the dynamic displacement characteristics during path switching. Calculation of displacement time derivatives involves converting displacement changes into velocity and acceleration parameters through mathematical differential operations. This can be achieved using numerical differentiation or polynomial fitting and derivation methods, such as calculating the rate of change of velocity at adjacent time points using the central difference method, thereby accurately describing the drone's motion state. Adaptive PID control algorithms dynamically adjust proportional, integral, and differential coefficients based on real-time motion parameters. Fuzzy logic or neural network optimization parameter adjustment strategies can be used, for example, to automatically adjust the weight coefficient of the integral term based on the rate of change of acceleration, thereby achieving dynamic optimization of control instructions.

[0074] Specifically, during the path segment switching process, a displacement change function is constructed through time series data acquisition. The displacement differences and time intervals between adjacent path segments are analyzed, and the instantaneous velocity and acceleration parameters are calculated. Combined with the attitude adjustment angle parameter, the motion state and attitude change are coupled and modeled, and the output of the control command is dynamically adjusted using an adaptive PID algorithm. For example, when a sudden acceleration change is detected, the algorithm automatically enhances the differential term to suppress overshoot; when the displacement deviation persists, the integral term is weighted to eliminate steady-state errors. This enables real-time optimization of flight control commands, ensuring a coordinated response between attitude adjustment and displacement changes. Thus, by combining displacement derivative calculation with an adaptive control algorithm, this solution achieves coordinated optimization of motion state and attitude parameters, effectively resolving the problem of control instability during dynamic flight.

[0075] Through the above technical solution, this application can calculate motion state parameters in real time and dynamically adjust control instructions when the drone performs dynamic path switching, ensuring that attitude adjustments and displacement changes are accurately matched. For example, in a scenario where a dynamic obstacle is avoided, the drone can quickly adjust its yaw angle based on real-time acceleration changes to avoid flight jitter or deviation caused by sudden path changes, thereby improving the stability and clarity of image acquisition.

[0076] The present application further proposes that step S3 includes the following steps:

[0077] Step S31: Matching the real-time position information of the UAV with the dynamic flight path, and calculating the deviation between the current position of the UAV and the target shooting point;

[0078] Step S32: adjusting the flight speed and direction of the drone in real time based on the deviation information to ensure that the drone can accurately reach the target shooting point;

[0079] Step S33: After the UAV reaches the target shooting point, the focus and exposure parameters of the camera are adjusted in real time according to the depth of field distribution map of the target area to ensure that the image clarity of each sub-area within the target area reaches a preset threshold;

[0080] Step S34: Verify the adjusted camera parameters. If the image clarity does not reach the preset threshold, readjust the camera parameters until the requirements are met.

[0081] Among them, real-time position information refers to the three-dimensional coordinate data of the drone obtained through a satellite positioning system or an inertial navigation system. Specifically, this can be achieved using GPS and IMU fusion positioning technology, and is used to track the drone's position offset relative to the preset path in real time. The dynamic flight path refers to the flight trajectory dynamically generated based on the depth of field characteristics of the target area. Specifically, this can be achieved using a path planning algorithm combined with a real-time obstacle avoidance module, and is used to guide the drone to execute the trajectory according to the optimal shooting distance. Deviation information refers to the coordinate difference between the drone's current position and the target shooting point. Specifically, this can be achieved using the Euclidean distance calculation method, and is used to quantify the degree of deviation from the flight path. The depth of field distribution map refers to a two-dimensional distribution map reflecting the optimal focus distance of each position in the target area. Specifically, this can be achieved using a multispectral imaging device combined with a depth estimation algorithm, and is used to guide the dynamic adjustment of camera parameters. The preset threshold refers to a quantitative indicator of image clarity. Specifically, it can be set using an image sharpness evaluation algorithm combined with the requirements of the application scenario, and is used to determine whether the image quality meets the standard.

[0082] Specifically, by acquiring the drone's position coordinates in real time and spatially matching them with the preset path, the lateral and longitudinal deviations of the flight trajectory are calculated. Based on the direction and magnitude of the deviation vector, the proportional-integral control algorithm is used to generate flight speed correction instructions and heading angle adjustment instructions, so that the drone converges to the target shooting point along the optimal path. When arriving at the shooting position, the camera is driven to automatically adjust the lens focal length to the optimal imaging state based on the optimal focus distance of the corresponding sub-area in the depth of field distribution map; at the same time, combined with the ambient light intensity, the optimal exposure parameter combination is calculated and set in real time through the light metering module. After the adjustment is completed, the clarity of the collected image is detected. If the preset sharpness threshold is not reached, the depth of field data is recalculated and the camera parameters are iteratively adjusted until the image quality requirements are met. Therefore, this solution forms a complete quality control closed loop through clarity detection and iterative adjustment.

[0083] Through the above technical solution, this application solves the problem of misaligned shooting points caused by positional offsets during dynamic flight, eliminates local imaging blur caused by fixed camera parameters, and enables autonomous acquisition of high-definition images in complex environments. Real-time path correction and parameter optimization reduce the frequency of manual intervention; a closed-loop verification mechanism ensures that the image quality of each shooting point meets engineering standards; and differentiated parameter settings adapt to the imaging requirements of different sub-depths of field characteristics in the target area.

[0084] The present application further proposes that step S4 includes the following steps:

[0085] Step S41: performing multi-scale feature extraction on the collected high-resolution image data to generate a multi-scale feature map of the target area;

[0086] Step S42: performing feature fusion on the image data of the target area based on the multi-scale feature map to generate an ultra-high-resolution panoramic image of the target area;

[0087] Step S43: performing geometric correction on the panoramic image using a feature matching algorithm to eliminate image distortion;

[0088] Step S44: generating a final high-definition image of the target area based on the geometrically corrected panoramic image.

[0089] Among them, multi-scale feature extraction refers to extracting feature information at different spatial resolutions from image data. Specifically, it can be achieved by using the multi-level feature fusion method of convolutional neural networks, and constructing a multi-scale feature map by integrating low-level detail features and high-level semantic features. Among them, feature fusion refers to spatial alignment and information integration of multi-scale feature maps. Specifically, it can be achieved by combining the channel attention mechanism with spatial pyramid pooling. The texture details and spatial continuity of panoramic images are enhanced by adaptively weighted fusion of features of different scales. Among them, the feature matching algorithm refers to a method for calculating geometric transformation parameters based on the correspondence between image feature points. Specifically, it can be achieved by using an improved random sampling consistency algorithm, and the interference of mismatched points on the homography matrix estimation is suppressed by a robust error function.

[0090] Specifically, the original image is decomposed into a pyramid through the multi-scale feature extraction module, and edge, texture and structural features are extracted at the coarse and fine granularity levels respectively. The feature fusion module performs channel dimension splicing and spatial dimension interpolation on the feature maps of different levels to achieve detail enhancement and resolution improvement. The geometric correction module uses the feature matching algorithm to calculate the homography transformation matrix between adjacent images, and eliminates the geometric deviation caused by lens distortion and projection deformation through iterative optimization. The image generation module performs global stitching and color equalization on the corrected multi-view images to output a geometrically consistent panoramic image. Therefore, this scheme realizes unsupervised registration through the multi-scale features of the image itself, uses feature fusion to enhance the self-similarity of the image, and combines the improved geometric correction algorithm to automatically eliminate distortion, and can complete high-precision image stitching without relying on external terrain data or manual intervention.

[0091] Through the above technical solution, this application solves the problem of low image registration efficiency caused by the lack of initial terrain information, improves the image self-matching ability through multi-scale feature fusion, and automatically eliminates lens distortion and projection deformation by combining the robust geometric correction algorithm, significantly reducing the workload of manual correction, and realizing the efficient generation of ultra-high-resolution panoramic images in complex scenes.

[0092] This application further proposes that the feature matching algorithm adopts an improved algorithm based on RANSAC, and its formula is as follows: ,in, Indicates the accumulation operation of all N matching point pairs. is the index of the matching point pair (from 1 to N), N is the total number of matching point pairs, Represents the homography matrix H for the Source image feature points The geometric transformation result, H represents the homography matrix, and Represent the coordinates of the matching point pairs, represents the Euclidean distance function, Represents a robust error function, which is used to reduce the influence of mismatched points. The algorithm achieves accurate geometric correction of panoramic images by iteratively optimizing the homography matrix H.

[0093] The homography matrix H is a parameter matrix that describes the geometric transformation relationship between images. It can be estimated by using the least squares method combined with the random sampling consistency algorithm to map the image coordinates at different perspectives to a unified plane. Refers to calculating the geometric deviation of the coordinates of the matching point pair after transformation, which can be achieved by taking the square root of the difference between the coordinates of the two points, and is used to quantify the spatial position difference of the matching point pair. It refers to a function that performs nonlinear mapping on the error, which can be implemented using Huber loss function or Tukey bi-weight function to reduce the interference of mismatched points on the optimization process.

[0094] Specifically, in the image feature matching process, a set of inliers is first selected from the initial matching point pairs through a random sampling consistency framework, and an initial estimate of the homography matrix H is constructed based on the inliers. Subsequently, the geometric deviations of all matching point pairs are calculated using the Euclidean distance function, and the deviations are weighted using a robust error function so that the contribution of mismatched points decays as the deviations increase. In the iterative optimization stage, the parameters of the homography matrix H are dynamically adjusted by minimizing the weighted total deviation until they converge to the optimal solution. This process unifies the suppression of mismatched points and the optimization of geometric transformation parameters into a computable mathematical problem through mathematical modeling, thereby eliminating the influence of outliers while retaining valid matching points.

[0095] Through the above technical solution, this application effectively reduces the interference of mismatched points on the accuracy of geometric correction, and can still maintain a stable correction effect in dynamic obstacles or complex texture scenes. At the same time, it reduces the need for manual reshooting and improves the degree of automation of the image processing process.

[0096] In summary, the present invention replaces the reliance on preset terrain data through multispectral imaging and dynamic depth of field analysis, enabling the UAV to autonomously plan paths in unknown terrains and improve environmental adaptability. Secondly, dynamic path planning combined with real-time parameter adjustment can cope with dynamic obstacles in complex environments, avoid manual reshooting, and improve work efficiency. Furthermore, multi-scale feature extraction and fusion technology, combined with an improved geometric correction algorithm, can generate high-precision panoramic images, eliminate distortion, and ensure image quality. In addition, strategies such as adaptive PID control ensure the smooth flight of the UAV, further improving image clarity and mission success rate. This solution forms a closed-loop control from data acquisition to image output, realizing automated, high-precision high-definition image acquisition, and has important application value in the fields of terrain mapping, disaster assessment, etc.

[0097] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, various modifications and variations of the present application are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A method for obtaining high-definition images by approaching a drone, characterized in that: The following steps are involved: Step S1: Collect high-resolution image data of the target area through the multispectral imaging equipment carried by the UAV, and simultaneously record the flight attitude parameters of the UAV; perform dynamic depth of field analysis on the collected image data to generate a depth of field distribution map of the target area; Step S2: Determine the optimal shooting distance for each sub-area within the target area based on the depth of field distribution map, and plan a dynamic flight path that meets the optimal shooting distance in combination with the real-time position information of the UAV; Step S2 includes the following steps: Step S21: Dividing the depth feature point set of the target area into sub-areas to generate multiple sub-areas of the target area; Step S22: determining the optimal shooting distance range for each sub-area based on the depth of field distribution map, and performing dynamic constraint analysis in combination with the current flight altitude and speed of the UAV to generate a dynamic flight path that meets the optimal shooting distance; Step S23: Optimize the dynamic flight path in real time to avoid obstacles that may exist in the path and ensure that the drone can smoothly transition to the next shooting point; Step S24: generating flight control instructions for the UAV according to the optimized dynamic flight path, and adjusting the flight attitude of the UAV in real time to ensure that the camera is always facing the target area; Step S24 includes the following steps: Step S241: decomposing the dynamic flight path into multiple path segments, and assigning corresponding flight attitude parameters to each path segment; Step S242: Decomposing the flight attitude parameters in each path segment into directional components to obtain attitude adjustment components of the UAV in the horizontal, vertical, and yaw directions; Step S243: performing an inverse solution on the attitude angle of the UAV based on the attitude adjustment components of the UAV in the horizontal, vertical, and yaw directions to obtain the attitude adjustment angle of the UAV between adjacent path segments; Step S244: generating flight control instructions for the UAV based on the attitude adjustment angle of the UAV between adjacent path segments and in combination with the dynamic flight path to ensure that the UAV can smoothly complete the path switching; Step S244 includes the following steps: Perform displacement change analysis on each path segment in the dynamic flight path in the time dimension to obtain the displacement change value of the UAV between adjacent path segments over time; The displacement time derivative is calculated based on the change in the displacement of the UAV between adjacent path segments over time to obtain the displacement velocity and acceleration of the UAV between adjacent path segments. The UAV is controlled based on its attitude adjustment angle, displacement velocity and acceleration between adjacent path segments and combined with an adaptive PID control algorithm to generate the final flight control instructions. Step S3: Controlling the UAV to fly automatically according to the dynamic flight path, and adjusting the focus and exposure parameters of the camera in real time during the flight to ensure that the image clarity of each sub-area within the target area reaches a preset threshold; Step S4: Multi-scale feature extraction and fusion are performed on the collected high-resolution image data to generate an ultra-high-resolution panoramic image of the target area, and the panoramic image is geometrically corrected through a feature matching algorithm to eliminate image distortion.

2. The method for obtaining high-definition images by approaching a drone according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: deploying a multispectral imaging device on the UAV to collect high-resolution image data of the target area; Step S12: Synchronously record the attitude parameters of the drone during flight, including pitch angle, roll angle, and yaw angle, through the drone's built-in inertial measurement unit; Step S13: performing pixel-level brightness equalization processing on the collected image data to enhance the contrast of the image data and reduce the influence of uneven illumination, thereby obtaining brightness-balanced image data; Step S14: performing dynamic depth of field analysis on the brightness balanced image data to calculate the depth of field value corresponding to each pixel point in the target area and generate a depth of field distribution map of the target area; Step S15: Input the depth distribution map into the improved deep learning model to extract the significant feature points and their descriptors in the target area, and perform non-maximum suppression on the significant feature points and their descriptors to generate a depth feature point set of the target area.

3. The method for obtaining high-definition images by approaching a drone according to claim 2, characterized in that: The non-maximum suppression of the salient feature points and their descriptors includes: The local texture characteristics of salient feature points and their descriptors are analyzed to analyze the neighborhood structure corresponding to the salient feature points in the image. The gradient amplitude histogram of the pixels in the neighborhood of the salient feature points is calculated to characterize the edge direction distribution characteristics of the salient feature points. For the descriptors, their numerical distribution laws and dimensional correlations in different bands are analyzed to obtain the local characteristic data set of salient feature points and descriptors. Based on the local characteristic dataset of salient feature points and descriptors, the salient feature points and their descriptors are preliminarily screened. The salient feature points and their descriptors are associated in the time dimension considering temporal continuity. The position changes of the salient feature points in adjacent frames and the similarity of the descriptors are analyzed. The salient feature points with abnormal position changes or the descriptors with descriptor similarity lower than a preset threshold are marked as objects to be excluded for screening and exclusion, thereby obtaining a set of salient feature points and descriptors after the preliminary screening. A multi-scale local extreme value region is defined for each point in the salient feature points and descriptor set after the initial screening, so as to determine the corresponding local extreme value region according to the pixel texture condition in its neighborhood, and obtain the multi-scale local extreme value region; Based on the multi-scale local extreme value region, non-maximum suppression is performed on the salient feature points and descriptor sets after the initial screening to generate the depth of field feature point set of the target area.

4. The method for obtaining high-definition images by approaching a drone according to claim 3, characterized in that: Specifically, the multi-scale local extreme value region is to set a 3x3 scale local extreme value region in the high texture complexity region to capture detail features, and to set a 7x7 scale local extreme value region in the low texture complexity region to consider overall features.

5. The method for obtaining high-definition images by approaching a drone according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: Matching the real-time position information of the UAV with the dynamic flight path, and calculating the deviation between the current position of the UAV and the target shooting point; Step S32: adjusting the flight speed and direction of the drone in real time based on the deviation information to ensure that the drone can accurately reach the target shooting point; Step S33: After the UAV reaches the target shooting point, the focus and exposure parameters of the camera are adjusted in real time according to the depth of field distribution map of the target area to ensure that the image clarity of each sub-area within the target area reaches a preset threshold; Step S34: Verify the adjusted camera parameters. If the image clarity does not reach the preset threshold, readjust the camera parameters until the requirements are met.

6. The method for obtaining high-definition images by approaching a drone according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: performing multi-scale feature extraction on the collected high-resolution image data to generate a multi-scale feature map of the target area; Step S42: performing feature fusion on the image data of the target area based on the multi-scale feature map to generate an ultra-high-resolution panoramic image of the target area; Step S43: performing geometric correction on the panoramic image using a feature matching algorithm to eliminate image distortion; Step S44: generating a final high-definition image of the target area based on the geometrically corrected panoramic image.

7. The method for obtaining high-definition images by approaching a drone according to claim 6, characterized in that: The feature matching algorithm adopts an improved algorithm based on RANSAC, and its formula is as follows: ,in, Indicates the accumulation operation of all N matching point pairs. is the index of the matching point pair, from 1 to N, where N is the total number of matching point pairs. Represents the homography matrix For the first Source image feature points The geometric transformation result, H represents the homography matrix, and Represent the coordinates of the matching point pairs, represents the Euclidean distance function, Represents a robust error function, which is used to reduce the influence of mismatched points. The algorithm iteratively optimizes the homography matrix H to achieve accurate geometric correction of panoramic images.

Images