Self-adaptive microwave radar navigation system of unmanned aerial vehicle

Abstract

The invention provides a self-adaptive microwave radar navigation system for an unmanned aerial vehicle, and the system comprises a route planning module which provides basic comprehensive navigation information; an image acquisition module; the feature extraction module is used for generating feature data; the target identification module is used for providing real-time data for the navigation information generation module; the image fusion module is used for generating a comprehensive image representing accurate environment information; the navigation information generation module controls the unmanned aerial vehicle to efficiently complete an inspection task; and the result output module is used for visually displaying the inspection task. According to the invention, by introducing the key area marking and identifying unit in the route planning module and utilizing the feature correlation value and the polygon intersection judgment formula, accurate identification and marking of the key area are realized, and the problem of inaccurate identification of the key area in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) flight inspection, and particularly to an adaptive microwave radar navigation system for UAVs. Background Art

[0002] The background of UAV inspection operation mainly stems from the limitations of traditional inspection methods and the rapid development and wide application of UAV technology. Limitations of traditional inspection methods: Traditional inspection methods, such as manual inspection of industrial parks or static monitoring equipment, can no longer meet the modern society's demand for efficient and intelligent inspection. Manual inspection is not only time-consuming and laborious, but also poses great safety hazards when facing complex terrains, adverse weather conditions or high-altitude operations. In addition, manual inspection is difficult to be comprehensive and detailed, and potential fault points are easily overlooked.

[0003] In recent years, with the continuous innovation and popularization of UAV technology, UAV inspection systems have emerged and developed rapidly. UAV inspection systems integrate advanced sensors, high-definition cameras and intelligent analysis systems on UAVs to achieve remote and non-contact inspection of specific areas or targets. However, due to problems such as static planning, weak image processing ability, non-intelligent data fusion and poor dynamic adaptability, UAVs have low inspection efficiency and high risks in complex industrial environments.

[0004] Traditional UAV inspection uses fixed flight routes and cannot dynamically adjust the flight density according to the importance of areas, resulting in insufficient inspection of key areas (such as dangerous goods storage areas). Especially in complex environments such as industrial parks, it is difficult to accurately identify and mark key areas, and it is impossible to dynamically identify the relevance between key areas and the surrounding environment. In industrial parks, there are complex spatial relationships between dangerous chemical storage areas and other areas, but existing technologies are difficult to identify these key areas through the correlation between business data points and sub-areas, resulting in the inability of inspection tasks to accurately focus on key areas.

[0005] At the same time, when correcting the distortion of UAV-transmitted images in existing technologies, calculations are usually directly based on the original image coordinate system, with high calculation complexity and low efficiency. For example, when processing a 1024×1024 image in existing technologies, directly calculating the radial distortion parameters requires a large amount of time and resources, affecting the overall processing speed and making it difficult to meet the requirements of real-time UAV inspection. Summary of the Invention

[0006] Aiming at the deficiencies of the existing technology, the purpose of the present invention is to provide an adaptive microwave radar navigation system for UAVs. To achieve the above purpose, the present invention is realized through the following technical solutions: The adaptive microwave radar navigation system for UAVs includes:

[0007] A flight route planning module, which provides basic and comprehensive navigation information for UAV flight missions;

[0008] An image acquisition module that acquires visible light images and radar data of the industrial park, preprocesses the acquired visible light image data, and improves the image quality after performing denoising and image enhancement operations;

[0009] A feature extraction module that extracts key feature data from the preprocessed visible light images. The key feature data at least includes building outlines, roads, and boundaries of key areas. Analyze the extracted key feature data to determine the boundaries of key areas and key location information in the industrial park, and then calculate the boundary pixel points of the key areas through an edge detection algorithm to generate feature data;

[0010] A target recognition module that performs target recognition on the generated feature data by constructing a deep learning algorithm. The recognized targets include buildings, vehicles, and personnel targets in the industrial park, and real-time tracking of the recognized targets is performed to generate the movement trajectories of the targets, providing real-time data for the navigation information generation module;

[0011] An image fusion module that fuses the preprocessed visible light image data and radar data to generate a comprehensive image representing accurate environmental information, calculates the fusion weight through the Kalman filter algorithm, and generates a more accurate environmental image;

[0012] A navigation information generation module that generates navigation information for the drone based on the fused comprehensive image to ensure that the drone avoids obstacles and completes the inspection task, thereby controlling the flight direction and speed of the drone to ensure that it can complete the inspection task safely and efficiently;

[0013] A result output module that visually displays the inspection task for the convenience of operators to monitor.

[0014] Compared with the prior art, the beneficial effects of the present invention are:

[0015] 1. By identifying and labeling key areas, the aim is to solve the problem in the prior art of inaccurate identification of key areas in complex environments such as industrial parks. The present invention proposes a method based on the feature correlation value and the polygon intersection judgment formula for labeling key areas of sub-regions, so as to ensure that the unmanned aerial vehicle (UAV) can accurately identify and focus on these areas. Specifically, by calculating the feature correlation value Cbc between business data points and sub-regions, it is possible to effectively identify the key areas in the industrial park that are associated with the business data points. For example, for a hazardous chemical storage area, by calculating its correlation value with the surrounding sub-regions, it is possible to accurately determine whether the area is a key area and label it. By accurately labeling key areas, the UAV can focus more on key areas when collecting images, thereby improving the accuracy and efficiency of image processing. In the subsequent image acquisition module, by performing distortion correction and edge detection on key areas, clearer and more accurate image information can be obtained, providing a comprehensive data basis for subsequent target recognition and navigation information generation;

[0016] 2. At the same time, by improving the edge detection algorithm and the image distortion correction method, the accuracy and efficiency of image processing are improved, ensuring that the UAV can obtain clear and accurate environmental images, thereby providing more reliable support for the navigation and inspection tasks of the UAV. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The disclosure of the present invention will be described with reference to the accompanying drawings. It should be understood that the drawings are only for illustrative purposes and are not intended to limit the scope of protection of the present invention. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0018] Figure 1 is a schematic diagram of the module framework of the UAV adaptive microwave radar navigation system proposed in an embodiment of the present invention;

[0019] Figure 2 is a schematic diagram of the process logic of the route planning module proposed in an embodiment of the present invention;

[0020] Figure 3 is a schematic diagram of the process logic of image acquisition and preprocessing proposed in an embodiment of the present invention;

[0021] Figure 4 is a schematic diagram of the process logic of feature extraction and target recognition proposed in an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0022] It is easy to understand that according to the technical solution of the present invention, without changing the essence of the present invention, those of ordinary skill in the art can propose various interchangeable structural forms and implementation manners. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solution of the present invention, and should not be regarded as all of the present invention or as a limitation or restriction on the technical solution of the present invention.

[0023] The present invention will be further described in detail below with reference to the drawings, but it is not a limitation to the present invention.

[0024] As Figures 1 - 4 shown, as an embodiment of the present invention, an unmanned aerial vehicle adaptive microwave radar navigation system is proposed, including:

[0025] A route planning module S1, which is used to provide basic and comprehensive navigation information for the unmanned aerial vehicle flight mission;

[0026] An image acquisition module S2, which obtains visible light images and radar data of the industrial park through a high-resolution camera and a microwave radar carried thereon, and preprocesses the collected image data. After performing denoising and image enhancement operations, the image quality is improved so that subsequent system modules can more accurately identify and analyze the situation in this area;

[0027] A feature extraction module S3, which extracts key feature data from the preprocessed image data. The key feature data at least includes building outlines, roads, and boundaries of key areas. Analyze the extracted key feature data to determine the boundaries and key position information of key areas in the industrial park. For example, in the warehouse area of the industrial park, the feature extraction module needs to identify the outline and entrance / exit positions of the warehouse, and then calculate the boundary pixel points of the warehouse outline through an edge detection algorithm, so as to generate accurate feature data for subsequent target recognition and navigation information generation;

[0028] A target recognition module S4, which performs target recognition on the extracted feature data by constructing a deep learning algorithm (such as a convolutional neural network). The recognized targets include buildings, vehicles, and personnel targets in the industrial park, and the recognized targets are tracked in real time to generate the movement trajectories of the targets. For example, in the factory building area (key area) of the industrial park, identify the production equipment and transport vehicles in the factory building, and calculate the displacement of the target between consecutive frames through the optical flow method to generate its movement trajectory and output it, so as to provide real-time data for the navigation information generation module;

[0029] Image fusion module S5 fuses the collected visible light images and radar data to generate a comprehensive image representing accurate environmental information. For example, in the hazardous chemical storage area of an industrial park, the image fusion module combines the chemical containers shown in the visible light image with the terrain obstacle information reflected in the radar data, calculates the fusion weight through the Kalman filtering algorithm, and generates a more accurate environmental image, thereby providing more accurate environmental information for the navigation information generation module;

[0030] Navigation information generation module S6 generates the navigation information of the drone based on the fused comprehensive image to ensure that the drone avoids obstacles and completes the task. For example, in the warehouse area of an industrial park, the navigation information generation module calculates the best path from the current point to the next checkpoint based on the warehouse outline shown in the comprehensive image, generates navigation information, and thus controls the flight direction and speed of the drone to ensure that it can complete the inspection task safely and efficiently;

[0031] Result output module S7 visually displays the inspection task. For example, it displays the fused image, the position of the target object, and the flight trajectory of the drone on the screen, facilitating the operator to monitor.

[0032] In an embodiment of the present invention, it can be understood that during the process of providing basic comprehensive navigation information, the route planning module needs to divide the industrial park into multiple sub - regions and plan parallel drone routes for each sub - region, so as to maximize the coverage of all regions of the industrial park by the drone. Specifically, during implementation, due to the complexity of the industrial park, in order to ensure the provision of basic comprehensive navigation information for the drone flight task, for key regions (such as factory buildings, warehouses, hazardous chemical storage areas, etc.), further route planning is required to obtain more detailed image information. For example, in an industrial park, if the factory building area is divided into a key sub - region, the route spacing needs to be reset to ensure that the dynamic situation of this region can be monitored more carefully and the drone can complete the inspection task efficiently. At the same time, according to the scale of the industrial park and the urgency of the task, multiple drones can also be assigned to perform the inspection task simultaneously, and different routes are assigned to each drone to avoid mutual interference.

[0033] Based on the above technical concept, the proposed route planning module S1 includes:

[0034] Map data acquisition unit, which acquires high - precision map data, and the process is as follows:

[0035] S1 - 1. In the industrial park, use lidar and satellite remote sensing technologies to collect the terrain data and building contour data of the industrial park, and obtain service data points p i ; S1 - 2. Replace the value of each service data point p i with its neighborhood Ni The median of the internal data points to effectively suppress noise: p i new = median(N i ), where p i new is the value of the data point after filtering; S1-3. Unify the coordinate systems of data points from different sources through projection transformation: Let the projection transformation matrix of the data points from the original coordinate system (x 1 , y 1 ) to the target coordinate system (x 2 , y 2 ) be M, then there is S1-4. Calculate the positions of each data point in the simulation grid of the navigation system: where row and col are the row number and column number of the data point in the y direction in the simulation grid respectively, and cell_size x , cell_size y are the side lengths of the grid cells in the x and y directions respectively, is floor function.

[0036] Sub-region preliminary division unit. Based on the density distribution of data points in the simulation grid, divide the industrial park into sub-regions. For example, for key areas such as factory buildings, warehouses, and hazardous chemical storage areas, achieve regional division by increasing the density threshold to ensure that the drones conduct fine inspections of key areas. The process is as follows:

[0037] S1-5. Set the inner radius e of the neighborhood N i and the minimum number of data points Min pts within this radius. For each grid cell Ci, calculate the number of points N e (Ci) in its neighborhood. If the number of points |N e (Ci)| ≥ Min pts , then determine Ci as a core cell, and |.| represents the number of elements in the set of points N e (Ci). At this time, expand e cells around the core cell Ci in all directions to count the number of cells in the neighborhood; and form a clustering cluster based on the area connected by the core cell Ci to complete the division of sub-regions.

[0038] Key area marking and recognition unit. Mark key areas for the sub-regions. The process is as follows:

[0039] S1-6. Based on the obtained business data points p i in the industrial park (such as the location information of factory buildings, warehouses, and hazardous chemical storage areas) and the divided sub-regions, through the following formula C bc calculate the business data point p iThe characteristic correlation value of the corresponding points of the polygon area (represented as a set of cells in the simulation grid) within the sub-area (also a set of cells). If this value exceeds the set threshold t, it is considered that the two areas are correlated near this point, indicating the existence of a key area. Then, the polygon intersection judgment formula Count(b∩c)>0 is used to judge the overlapping situation of the two areas. If the condition is met, it means that the sub-area contains the key area and is marked. In the formula, count is a function for counting the number of cells in the intersection of the two sets of cells, b is the polygon area in the business data point p i in, and c is the sub-area, and the formula C bc is expressed as follows:

[0040]

[0041] In the formula, x bc is the distance between a data point in the polygon area b and a data point in the sub-area c in the simulation grid. In the simulation grid, this distance is measured in the number of cells and is used to measure the spatial relationship between points in the two areas. The value of it reflects the degree of proximity between the two points. The smaller the distance, the closer the two areas are near this point; s is a parameter related to the characteristic intensity of the key area in the business data. For example, for the storage area of hazardous chemicals, s represents the quantified value of its hazard level. The larger the s value, the higher the hazard level of the area and the greater the importance weight in the specific identification of key areas; C s 、C x 、C z are the scale parameters related to s, x bc 、Z c and H b respectively, which are used to adjust the influence degree of each part in the formula C bc . Their values are set according to the actual situation of the industrial park and the characteristics of the business data. For example, for an area with complex terrain, the value of C z needs to be adjusted accordingly to adapt to the influence of the terrain on the area division; Z c is the attribute value of a data point in the sub-area, such as the altitude of this data point, which is used to consider the characteristics of the sub-area in the vertical direction; H b is the attribute value corresponding to a data point in the polygon area b of the business data point p i , such as the height range of the factory building in the vertical direction; μ is a normalization constant, which is used to ensure that the value of x bc is within a reasonable range to facilitate the combination with other judgment conditions; n is the function influence parameter. By adjusting the value of n, the sensitivity of the formula C bc to different distances and attribute values is changed to adapt to the identification requirements of different types of key areas.

[0042] Based on the above technical concept, it should be noted that the purpose of the present invention is to solve the problem of inaccurate identification of key areas in complex environments such as industrial parks by identifying and annotating key areas. In traditional drone inspections, due to the lack of accurate identification and annotation of key areas (such as hazardous chemical storage areas, factory buildings, warehouses, etc.), drones are unable to effectively focus on key areas during mission execution, thus affecting inspection efficiency and safety. In addition, it is difficult for the prior art to identify these key areas through the correlation between business data points and sub-areas, resulting in the inspection task being unable to accurately focus on key areas. The present invention proposes a method based on the feature correlation value and the polygon intersection judgment formula for annotating key areas of sub-areas to ensure that drones can accurately identify and focus on these areas. Specifically, by calculating the feature correlation value Cbc between business data points and sub-areas, key areas associated with business data points in the industrial park can be effectively identified. For example, for a hazardous chemical storage area, by calculating its correlation value with surrounding sub-areas, it can be accurately determined whether the area is a key area and annotated. By accurately annotating key areas, drones can focus more on key areas when collecting images, thereby improving the accuracy and efficiency of image processing. In the subsequent image acquisition module, by performing distortion correction and edge detection on key areas, clearer and more accurate image information can be obtained, providing a comprehensive data basis for subsequent target recognition and navigation information generation.

[0043] A flight path spacing calculation and generation unit, for each sub-area, calculates the flight path spacing according to the regional terrain slope, the drone flight altitude, the camera field of view angle, and the lens distortion parameters, and within the simulated grid corresponding to each sub-area, based on the calculated flight path spacing, calculates the cost function of each grid cell data point by using the Manhattan distance, and selects the path with the minimum cost as the initial flight path data; at the same time, for key areas, optimizes the initial flight path data according to its boundary and obstacle information, and the optimization method is to use Bezier curve interpolation to smooth each initial flight path data to generate a smooth curve, and then reconstruct the flight path, and use the smooth curve as the new flight path data to make the drone flight path more conform to the key area for subsequent acquisition of detailed image information.

[0044] In an embodiment of the present invention, the image acquisition module S2 includes:

[0045] An image acquisition unit, based on the new flight path data, real-time acquires visible light images during the flight of the drone to ensure that clear and accurate environmental images can be obtained;

[0046] A data preprocessing unit, preliminarily processes the acquired visible light images to improve the image quality, and the process is as follows:

[0047] S2-1. Perform noise removal. Using filtering algorithms such as median filtering and Gaussian filtering, remove the noise in the visible light image to reduce the interference of noise on subsequent processing. Specifically, during implementation, based on the key area range obtained by the key area annotation and recognition unit, for each pixel point (i, j) in the image within the key area, assume its neighborhood window size is n (usually n is an odd number, such as 3*3, 5*5), then according to the service data point p i Determine the key area, and obtain the neighborhood pixel value set as Sort all the pixel values in this neighborhood by size to obtain

[0048] Take the middle value as the pixel value I of the pixel point (i, j) after filtering filtered (i, j), In the formula,[[]]END]] Is the median position of the sorted pixel value set;

[0049] S2-2. Perform visible light image correction. Perform distortion correction on the image to compensate for the image distortion caused by lens distortion factors, making the image closer to the real scene. The specific implementation steps include:

[0050] S2-21. In the stage of obtaining high-precision map data, determine the pre-calibrated radial distortion coefficients (r1, r2, r3, r4, r5) and tangential distortion coefficients (q1, q2, q3, q4, q5) of the extraction camera through a preset calibration method, as the basic data for visible light image distortion correction calculation. At the same time, based on the navigation system simulation grid mapping in the map data acquisition unit, obtain the position information of the visible light image in the target coordinate system, so as to determine the center coordinates (c x , c y ) of the visible light image within the key area. Calculate the distance d from each pixel point (i, j) in the image within the key area to the center of the visible light image through the distance calculation formula

[0051] ​Based on the above technical concept, it should be noted that in the UAV adaptive microwave radar navigation system, distortion correction is a key step to improve image quality and navigation accuracy. Generally, UAV navigation systems have high requirements for image accuracy, especially in key areas such as factory buildings and warehouses, where precise distortion correction is needed to ensure image quality. The accurate calculation of radial distortion and tangential distortion parameters directly affects the accuracy of the image, and any small error may lead to navigation errors. For example, when a UAV performs an inspection task in an industrial park, the following areas need to be monitored key points: Factory buildings: Due to frequent production activities and dynamic changes, high-precision images are required to monitor production conditions; Warehouses: The stacking and movement of goods cause image deformation, and precise distortion correction is needed to obtain accurate inventory information; Hazardous chemical storage areas: High safety requirements, and real-time and accurate images are needed to monitor the storage status. However, in the existing technology, the correction calculations of radial distortion and tangential distortion are usually directly based on pixel points in the original image coordinate system. The defects of this method are mainly reflected in the following aspects: High computational complexity: Directly performing distortion correction calculations in the original coordinate system involves complex polynomials, increasing the computational burden; Low efficiency: Due to high computational complexity, the processing speed is slow, affecting the real-time performance of the overall system; Difficult to manage: The lack of intermediate variables makes the calculation process difficult to manage and optimize, and it is difficult to adapt to the requirements of different scenarios.

[0052] In order to achieve efficient and concise distortion correction calculations, by introducing an intermediate variable, the calculation of the radial distortion correction part of the image within the key area is completed, thereby simplifying the complex radial distortion calculation into a certain calculation, reducing the computational complexity. For example, in the existing technology, when processing a 1024×1024 image, directly calculating the radial distortion parameters requires a large amount of time and resources, affecting the overall processing speed. However, after introducing the intermediate variable for calculation, the radial distortion correction result can be quickly obtained, thereby reducing the calculation time and improving the processing speed. The process is as follows:

[0053] S2-22. After completing the calculation of the distance from the pixel point to the center of the visible light image, perform radial distortion correction and tangential distortion correction, and obtain the corrected pixel gray value:

[0054] First, use the intermediate variable B, B = 1 + r 1 d 2 + r 2 d 4 + r 3 d 6 + r 4 d 8 + r 5 d 10 (5), perform radial distortion correction on the visible light image to correct the deformation generated by the lens in the radial direction, and then obtain the coordinates of the pixel points of the corrected visible light image (iradial , j radial );

[0055] Secondly, define intermediate variable A for tangential distortion correction i and A j :

[0056] A i = i·B + 2q 1 ·(i·B)·(j·B) + q 2 (d 2 + 2(i·B)^2 ) + … + q 5 (d 8 + 5(i·B) 8 )

[0057] A j = j·B + q 1 (d 2 + 2(j·B) 2 ) + q 2 (d 4 + 3(j·B) 4 ) + … + q 5 (d 10 + 6(j·B) 10 ) (6), using the tangential distortion coefficients (q1, q2, q3, q4, q5) and the coordinates (i radial , j radial ), perform tangential distortion correction to further eliminate image distortion caused by lens tangential deformation, and finally obtain the coordinates (i', j') of the visible light image pixels after complete correction. Among them, if the corrected pixel coordinates (i', j') are not integer coordinates in the original image, the gray value at this position needs to be obtained based on the bilinear interpolation principle:

[0058] I(i′, j′) = (1 - u)(1 - v)I(i 0 , j 0 ) + u(1 - v)I(i 1 , j 0 ) + (1 - u)vI(i 0 , j 1 ) + uvI(i 1 , j 1 ) (7),

[0059] In the formula, u and v are the weight parameters in the bilinear interpolation algorithm,

[0060] Finally, perform quality verification on the corrected visible light image of the key area. By comparing the image features before and after correction or comparing the corrected image with known real - scene information, check whether there are still obvious distortions or distortions. If it is found that the correction effect does not meet the expectations, adjust the camera parameters according to the actual situation and re - execute the distortion correction process to obtain a more accurate and clear visible light image.

[0061] In an embodiment of the present invention, the feature extraction module S3 includes:

[0062] An edge detection unit, using an edge detection algorithm, such as the Canny edge detection algorithm, to detect the edge information in the visible light image and extract the contours of the objects in the key area. Its specific implementation steps include:

[0063] S3 - 1. Calculate the gradient of the pre - processed visible light image to detect the change in pixel intensity in the image, and use the Sobel operator to calculate the gradients Gx and Gy of the visible light image in the horizontal and vertical directions respectively to obtain the gradient magnitude G of the visible light image.

[0064] S3 - 2. Perform non - maximum suppression on the calculated gradient magnitude image to refine the edges of the object contours in the key area, remove false edges, and at the same time, according to the gradient direction Compare at the gradient direction of each pixel point (i', j'), retain the local maximum value, suppress other non - maximum points, and then use the double - threshold method to detect and connect the edges. Among them, set two thresholds T h and T l respectively used to detect the strong edges of the object contours in the key area and the weak edges of the object contours in the key area. Then, for the gradient magnitude G, if G > T h , then the pixel point (i', j') is determined to be a strong edge point and is directly retained. If T l ≤G≤T h , then the pixel point (i', j') is determined to be a weak edge point, and it is necessary to further determine whether it is connected to a strong edge point. If it is connected, it is retained; otherwise, it is suppressed. If G < T l , then directly suppress the (i', j').

[0065] Based on the above technical details, it is necessary to understand that edge detection is a basic technology in image processing, which is used to identify the contours and boundaries of objects in images. This information is crucial for subsequent image analysis and understanding. In actual applications, due to factors such as noise, illumination changes, and image resolution, edge detection will produce errors, resulting in inaccurate or incomplete extracted edge information. When drones perform inspection tasks in industrial parks, in factory workshops, high-precision images are required to monitor production conditions. Due to the existence of edge detection errors, drones cannot accurately identify the boundaries of equipment on the production line, thereby affecting the monitoring and analysis of the production process. In addition, edge detection errors will also make it difficult for drones to accurately identify the contours and positions of goods, thereby affecting inventory management and logistics scheduling. For hazardous chemical storage areas, edge detection errors lead to inaccurate identification of storage containers, increasing safety risks.

[0066] At this point, it is necessary to reduce the edge detection error during the edge detection process, that is, after completing the edge extraction, the errors that may occur in this process (image segmentation and stitching process) can be effectively improved through weighted fusion algorithm and multi-threaded parallel processing to reduce the error, so as to better meet the rapid response requirements of drones in dynamic environments. The specific process is:

[0067] S3-3. For the overlapping areas between adjacent image blocks during the image segmentation process, the width is set to N pixels. Then, for each overlapping area, a linearly increasing weight function w(x, y) is defined to ensure the smoothness of the edge fusion of the object contour in the key area: x, y are the position index of each pixel (i', j') in the overlapping area. end ,y end ) and (x start ,y start ) represent the end and start position indexes of the overlapping area respectively;

[0068] S3-4. For each pixel point (i', j') in the overlapping area, the corresponding weight value is calculated according to its position index x, y, and the edge values ​​of adjacent image blocks are fused using the weighted average method: In the formula, E merged (x, y) is the merged edge value, E i (x, y) is the edge value of the i-th image block at position index (x, y), w i (x, y) is the weight value of the i-th image block at the position index (x, y), M is the number of image blocks involved in the fusion, where for non-overlapping areas, the original edge value is used directly without weighting processing.

[0069] S3-5. Compare the weighted fused edge image with the original edge image to check whether there is obvious error or distortion. If it is found that the edge error in some areas is still large, adjust the weight distribution according to the actual situation, such as increasing the weight of the central area or adjusting the slope of the linear increase. At this point, through accurate edge information, the drone can more accurately identify and analyze various objects and situations in the industrial park, thereby ensuring the safe and efficient operation of the industrial park.

[0070] The feature extraction module S3 also includes:

[0071] The image texture feature extraction unit is used to analyze the texture information of the visible light image and extract texture features, such as texture direction and texture density, so as to distinguish different terrains and objects. The specific implementation steps include:

[0072] S3-6, select a local window W of appropriate size, W = n*n pixels, slide it on the edge image processed by step S3-5, and for each area covered by window W, set its window center coordinates to (x w ,y w ), then the coordinate range of the pixel points in the window is: At this time, calculate the mean gradient amplitude in the window In one embodiment of the present invention, if there are N non-overlapping windows on the image, the window center coordinates are (x wi ,y wi ), i = 1, 2, ..., N, calculate the average value of the gradient amplitude in all windows Then, the standard deviation of the mean gradient amplitude in all windows is calculated As a measure of the texture density of the texture information characterizing the edge image, it is used to obtain the difference between the degree of texture change in different local areas of the edge image.

[0073]

[0074] Based on the above concept, it can be understood that in different local areas of the edge image, the texture can be regarded as a pattern of pixels with similar characteristics that are repeated or arranged regularly. For example, in the image of a brick wall, the arrangement pattern of bricks constitutes a texture; in the image of a grass field, the distribution of grass leaves is also a texture. A high texture density means that the texture elements are distributed more closely, more complexly, and more varied in the image; a low texture density means that the texture elements are distributed more sparsely, simply, and less varied. Therefore, the edge image is divided into multiple local windows W, and the mean gradient amplitude in each window is calculated. It means that the average gradient magnitude of each window represents the average degree of texture change in different local regions of the edge image. In specific implementation, for example, the average gradient magnitude within a certain window is relatively large, indicating that the texture change in this local region is relatively intense and contains rich texture information; conversely, if the average value is small, it means that the texture in this local region is relatively smooth and there is less texture information.

[0075] In an embodiment of the present invention, the target recognition module S4 includes:

[0076] A target recognition unit that uses an image recognition algorithm, such as a convolutional neural network (CNN) in deep learning, to recognize and classify target objects in an image. Its specific implementation steps include:

[0077] First, prepare the visible light image after distortion correction, edge detection, and texture feature extraction. Collect image data related to the target object in an order of magnitude and perform annotation to construct a training data set;

[0078] Secondly, use the convolutional neural network CNN architecture in deep learning, such as ResNet and VGG. Input the training data set into the CNN model for training, and during this training process, continuously adjust the parameters of the model to minimize the loss function so that the model can learn the features of the target object; in an embodiment of the present invention, the aforementioned CNN model is composed of a convolutional layer, a pooling layer, and a fully connected layer.

[0079] Among them, for the convolutional layer, let the input feature map of the l-th convolutional layer be X l-1 , the convolutional kernel be K l , the bias be b l , and the stride used for the convolutional operation be s and the padding be p. Then, the output feature map X l of the l-th convolutional layer is obtained. The calculation formula is:

[0080]

[0081] In the formula, (i, j, k) are the coordinates of the output feature map, M*N is the size of the convolutional kernel, C l-1 is the number of channels of the input feature map. At the same time, after the convolutional operation, the activation function ReLU is used to introduce non-linearity.

[0082]

[0083] For the pooling layer, taking max pooling as an example, let the pooling window size be P*Q and the stride be s p , then the output feature map X l of the l-th pooling layer is obtained. The calculation formula is:

[0084]

[0085] For the fully connected layer, let the input vector of the L-th fully connected layer be x, the weight matrix be W L , and the bias vector be b L . Then the output vector y of this fully connected layer is: y = W L x + b L .

[0086] Again, after the training is completed, input the visible light image to be recognized into the trained CNN model, identify and classify the target objects in the image, output the category information of the target objects, and determine whether the objects in the visible light image are production equipment, goods, or dangerous chemical storage containers in the industrial park.

[0087] In an embodiment of the present invention, the following data including the path of the visible light image file, the object category in the visible light image, and the annotation information of the visible light image in Table 1 are used as examples to demonstrate the local process of training and testing a convolutional neural network (CNN) model:

[0088]

[0089] In Table 1, the image file path: points to the complete path where the image file is stored; object category: the category of the main object in the visible light image, such as "production equipment", "goods", "dangerous chemical storage container"; bounding box coordinates: the coordinates of the bounding box annotating the position of the object in the visible light image, in the format of (x_min, y_min, x_max, y_max), where (x_min, y_min) are the coordinates of the upper left corner of the bounding box, and (x_max, y_max) are the coordinates of the lower right corner of the bounding box; image size: the width and height of the image, in pixels.

[0090] The divided dataset is shown in Table 2 below:

[0091]

[0092] The divided validation set is shown in Table 3 below:

[0093]

[0094] The divided test set is shown in Table 4 below:

[0095]

[0096] Then, the process of training and testing a Convolutional Neural Network (CNN) model is as follows: Load the visible light image file to be recognized, adjust the image to the input size used during the training of the CNN model, such as adjusting it to 224x224 pixels, and normalize the image according to the normalization method used during training, such as normalizing it using the same mean and standard deviation. Load the already trained CNN model and set the evaluation mode. Input the preprocessed image into the CNN model, extract the image features through the convolutional layer and pooling layer in sequence, process it through the fully connected layer, and finally classify the features through the fully connected layer. Apply the Softmax function after the last fully connected layer to convert the output into a probability distribution, and select the category with the highest probability as the prediction result. Display the recognition result (object category) and confidence level (probability value) to the user, and at the same time record the recognition result in a log file or database for subsequent analysis and auditing.

[0097] In an embodiment of the present invention, the specific implementation steps for the image fusion module S5 to provide more accurate environmental information for the navigation information generation module include:

[0098] First, normalize the visible light image I v (x, y) and radar data R(m, n) that have undergone preliminary processing (such as distortion correction, edge detection, and texture feature extraction) respectively to unify the data ranges of the two;

[0099] Second, establish the state model x k+1 = F k * x k + B k * u k + w k and the observation model z vk = H vk * x k + v vk 、z rk = H rk * x k + v rk for Kalman filtering, where the state model describes the motion law of the target object, and the observation model represents the position information of the target object obtained through the image;

[0100] Third, in each frame of the image during the Kalman filtering iteration, use the Kalman filtering algorithm to predict the state of the target object at the current moment based on the state at the previous moment, and then update the predicted state by combining the observation information of the target object in the current frame image to obtain the accurate position of the target object in the current frame; and update the covariance estimation to determine the fusion weight a;

[0101] Fourth, according to the formula I fused (x, y) = a * I vnorm(x, y) + (1 - a) * R norm (x, y) fuses the normalized visible light image and radar data to generate a composite image, where a is determined according to the Kalman filter result;

[0102] Finally, post-process the composite image through the formula I final (x, y) = I fused (x, y) * (max final - min final ) + min final Restore the pixel value to the original range to obtain the I final (x, y) as the environmental information image that provides more accurate environmental information for the navigation information generation module.

[0103] In an embodiment of the present invention, the navigation information generation module S6 includes:

[0104] A position estimation unit that estimates the current position of the UAV according to the features in the environmental information image and the position information of the target object. In specific implementation, first use the SIFT or SURF algorithm to extract feature points and descriptors from the fused environmental information image, and match them with the features in the map database pre-established by the system. Then, use the triangulation principle to calculate the current position of the UAV using the matching feature pairs and their corresponding position information.

[0105] A heading calculation unit that calculates the heading of the UAV according to the position of the target object and the current position of the UAV, and provides guidance for the flight direction of the UAV. In specific implementation, first identify the target object, such as the next checkpoint, from the environmental information image, and obtain its position coordinates xtarget, ytarget in the image coordinate system; secondly, convert this coordinate to the position in the actual geographical coordinate system through the pre-calibrated mapping relationship between the image and the geographical coordinates, and then obtain the current position xdrone, ydrone of the UAV based on the position estimation unit; thirdly, according to the position coordinates of the target object and the UAV, use trigonometric functions to calculate the heading angle θ of the UAV relative to the target object, θ = arctan2(ytarget - ydrone, xtarget - xdrone), where arctan2 is the four-quadrant arctangent function, which can determine the quadrant where the heading angle is located according to the positive and negative of the coordinate difference; finally, appropriately adjust the calculated heading angle to make it meet the flight control requirements of the UAV, and output the finally determined heading angle as part of the navigation information to provide guidance for the flight direction of the UAV.

[0106] Obstacle avoidance information generation unit: By analyzing the obstacle information in the environmental information image, it generates obstacle avoidance instructions to help the UAV avoid obstacles. Specifically, in implementation, first, based on the current position, heading, and speed of the UAV, combined with the position and range of the obstacle, the distance between the predicted flight path of the UAV and the obstacle boundary is calculated to evaluate the collision risk. If this distance is less than the safety threshold, it is considered that there is a collision risk. Secondly, when it is evaluated that there is a collision risk, corresponding obstacle avoidance strategies are formulated. The obstacle avoidance strategies at least include changing the flight direction and adjusting the flight altitude. For example, if an obstacle is detected ahead, the new heading after deviating a certain angle to the left or right can be calculated to avoid the obstacle. Finally, the formulated obstacle avoidance strategies are converted into specific obstacle avoidance instructions, such as the angle of changing the heading and the value of adjusting the flight altitude. The obstacle avoidance instructions are combined with the heading instructions output by the heading calculation unit to generate complete navigation information, controlling the UAV to avoid obstacles and ensuring flight safety.

[0107] In an embodiment of the present invention, the result output module S7 includes:

[0108] Navigation information sending unit, which converts the generated navigation information into flight control instructions and then sends them to the flight control system of the UAV based on Wi-Fi and Bluetooth wireless communication methods;

[0109] Visualization display unit, which visually displays the result of the preprocessed visible light image, such as displaying the fused image, the position and trajectory of the target object on the screen, facilitating the operator to monitor.

[0110] Taking Python as an example, if the serial communication protocol is used, the code example for converting and sending navigation instructions is as follows:

[0111] # Convert natural language navigation instructions into the UAV control protocol format

[0112] nav instruction = "Heading angle 45 degrees, move forward 100 meters"

[0113] flight_control_command = f"SET_HEADING,45\nMOVE,100"

[0114] # Send the instruction to the UAV (pseudo-code)

[0115] send_command_to_drone(flight_control_command).

[0116] The code example for image loading and visualization is as follows:

[0117] # Load the preprocessed image

[0118] image_path = "preprocessed_image.jpg"

[0119] image = cv2.imread(image_path).

[0120] An example code for target trajectory and position annotation is as follows:

[0121] # Define the target position and trajectory (example data)

[0122] target_position = (250, 150)

[0123] target_trajectory = [(100, 100), (200, 200), (250, 150)]

[0124] # Draw green trajectory points

[0125] for pos in target_trajectory:

[0126] cv2.circle(image, pos, 5, (0, 255, 0), -1) # Green point, radius 5, filled

[0127] # Draw the current position of the red target

[0128] cv2.circle(image, target_position, 10, (0, 0, 255), -1) # Red point, radius 10, filled.

[0129] An example code for image display and interaction is as follows:

[0130] # Display the annotated image

[0131] cv2.imshow('Visualization', image)

[0132] cv2.waitKey(0) # Wait for the user to press a key (0 means infinite wait)

[0133] cv2.destroyAllWindows() # Close all OpenCV windows.

[0134] The technical scope of the present invention is not limited to the content described above. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. UAV adaptive microwave radar navigation system, characterized by: include: The route planning module provides basic and comprehensive navigation information for UAV flight missions; The image acquisition module acquires visible light images and radar data of the industrial park, and pre-processes the acquired visible light image data, performs denoising and image enhancement operations, and improves the image quality; A feature extraction module extracts key feature data from the preprocessed visible light image, wherein the key feature data includes at least the outline of buildings, roads, and the boundaries of key areas, analyzes the extracted key feature data to determine the boundaries and key location information of the key areas in the industrial park, and then calculates the boundary pixel points of the key areas through an edge detection algorithm to generate feature data; The target recognition module performs target recognition on the generated feature data by constructing a deep learning algorithm. The targets to be recognized include buildings, vehicles, and personnel targets in the industrial park. The recognized targets are tracked in real time to generate the target's motion trajectory to provide real-time data for the navigation information generation module. An image fusion module fuses the pre-processed visible light image data and radar data to generate a comprehensive image representing accurate environmental information, and calculates the fusion weight through a Kalman filter algorithm to generate a more accurate environmental image; The navigation information generation module generates the navigation information of the UAV based on the fused comprehensive image to ensure that the UAV avoids obstacles and completes the inspection task, thereby controlling the flight direction and speed of the UAV to ensure that it can complete the inspection task safely and efficiently; The result output module visualizes the inspection tasks to facilitate monitoring by operators.

2. The UAV adaptive microwave radar navigation system according to claim 1 is characterized by: In the process of providing basic comprehensive navigation information, the route planning module needs to divide the industrial park into multiple sub-areas and plan parallel routes for drones for each sub-area, so as to maximize the drone's comprehensive coverage of various areas in the park. The route planning module includes: The map data acquisition unit is used to obtain high-precision map data. The process is as follows: S1-1. In the industrial park, laser radar and satellite remote sensing technology are used to collect the terrain data and building outline data of the industrial park to obtain business data points p i ; S1-2, each business data point p i The value is replaced by its neighbor N i The median of the inner data points is used to effectively suppress noise; S1-3, unify the coordinate systems of data points from different sources through projection transformation; S1-4, calculating the position of each data point in the navigation system simulation grid; The preliminary regional division unit divides the industrial park into sub-regions based on the density distribution of data points in the simulation grid, and the process is as follows: S1-5, let the neighborhood N i Inner radius e and the minimum number of data points within this radius Min pts , for each grid cell Ci, calculate the number of points N in its neighborhood e (Ci), if the number of points in the neighborhood |N e (Ci)|≥Min pts , then Ci is determined to be a core cell, and |.| represents the number of points N e (Ci) The number of elements in the set. At this time, the core cell Ci is taken as the center and expanded to the surrounding e cells to count the number of cells in the neighborhood; and a cluster is formed based on the area connected by the core cell Ci to complete the division of the sub-area.

3. The UAV adaptive microwave radar navigation system according to claim 2 is characterized by: The route planning module further includes: a key area marking and identification unit, which is used to mark the key areas of the sub-areas, and the process is as follows: S1-6, based on the business data points p in the industrial park i And the divided sub-areas are calculated by the following formula C bc Calculate business data point p i The characteristic correlation value of the polygonal area and the corresponding points in the sub-area, formula C bc It is expressed as follows: If the feature correlation value exceeds the set threshold t, it is considered that the two areas are correlated near the point, indicating that there is a key area; Then, the polygon intersection judgment formula Count(b∩c)>0 is used to judge the overlap of the two areas. If the condition of Count(b∩c)>0 is met, it means that the sub-area contains the key area and it is marked. In the formula, count is a function used to count the number of cells in the intersection of two cell sets, and b is the business data point p i The polygonal area in the figure, c is the sub-area, x bc is the distance between a data point in polygonal area b and a data point in sub-area c in the simulated grid. In the simulated grid, this distance is measured in units of cells and is used to measure the spatial relationship between points in two areas. Its value reflects the proximity between the two points. The smaller the distance, the closer the two areas are near the point. s is a parameter related to the characteristic intensity of key areas in business data. The larger the s value, the higher the danger level of the area and the greater the importance weight in the specific key area identification. C s , C x , C z are respectively related to s and x bc , Z c and H b The relevant scale parameter is used to adjust the formula C bc The impact degree of each part is set according to the actual situation of the industrial park and the characteristics of business data; c is the attribute value of a data point in the sub-region, which is used to examine the characteristics of the sub-region in the vertical direction; H b is the business data point p i The attribute value corresponding to a data point in the polygonal area b; μ is a normalization constant used to ensure that x bc The value of is within a reasonable range; n is the influencing parameter of the function.

4. The UAV adaptive microwave radar navigation system according to claim 1 is characterized by: The image acquisition module includes: The image acquisition unit collects visible light images of the drone in real time during flight to ensure that clear and accurate environmental images can be obtained; The data preprocessing unit performs preliminary preprocessing on the collected visible light images to improve the image quality. The process is as follows: S2-1, perform noise removal, use filtering algorithm to remove noise in visible light image, and reduce the interference of noise on subsequent processing: it performs the following operations: based on the key area range obtained by the key area labeling recognition unit, for each pixel point (i, j) in the visible light image in the key area, set its neighborhood window size to n, then according to the business data point p i Determine the key area, obtain the neighborhood pixel value set, sort all the pixel values ​​in the neighborhood pixel value set by size, and take the middle value as the pixel point (i, j) as the pixel value after filtering; S2-2, perform visible light image correction, perform distortion correction on the image, compensate for the image distortion caused by the system camera lens distortion factor, so as to make the image closer to the real scene, and perform the following operations: S2-21. In the stage of obtaining high-precision map data, the radial distortion coefficients (r1, r2, r3, r4, r5) and tangential distortion coefficients (q1, q2, q3, q4, q5) pre-calibrated by the camera are determined and extracted through a preset calibration method to serve as the basic data for the visible light image distortion correction calculation. At the same time, based on the navigation system simulation grid mapping in the map data acquisition unit, the position information of the visible light image in the coordinate system is obtained, so as to determine the center coordinates (c x ,c y ), through the distance calculation formula Calculate the distance d from each pixel point (i, j) in the visible light image within the key area to the center of the visible light image; S2-22, perform radial distortion correction and tangential distortion correction, and obtain the corrected pixel grayscale value: First, use the intermediate variable B, B = 1 + r1d 2 +r2d 4 +r3d 6 +r4d 8 +r5d 10 The radial distortion correction is performed on the visible light image to correct the deformation of the camera lens in the radial direction, and then the pixel coordinates of the corrected visible light image are obtained (i radial ,j radial ); Secondly, define the intermediate variable A for tangential distortion correction i and A j : Using the tangential distortion coefficients (q1, q2, q3, q4, q5) and the coordinates after radial distortion correction (i radial ,j radial ), perform tangential distortion correction, further eliminate the image distortion caused by the tangential deformation of the lens, and obtain the fully corrected visible light image pixel coordinates (i', j'). If the corrected pixel coordinates (i', j') are not integer coordinates in the original image, it is necessary to obtain the grayscale value of the position based on the bilinear interpolation principle: I(i′,j′)=(1-u)(1-v)I(i0,j0)+u(1-v)I(i1,j0)+(1-u)vI(i0,j1)+uvI(i1,j1), where u and v are weight parameters in the bilinear interpolation algorithm. Finally, the quality of the corrected visible light image of the key area is verified by comparing the image features before and after correction or comparing the corrected image with known real scene information to check whether there is still obvious distortion. If it is found that the correction effect does not meet the expectations, the camera parameters are adjusted according to the actual situation and the distortion correction process is re-executed to obtain a clear visible light image.

5. The UAV adaptive microwave radar navigation system according to claim 1 is characterized by: The feature extraction module comprises: The edge detection unit uses an edge detection algorithm to monitor the edge information in the acquired clear visible light image and extract the contours of objects in the key area. The process is as follows: S3-1, performing gradient calculation on the preprocessed visible light image to detect changes in pixel intensity in the image, and using the Sobel operator to calculate the gradients Gx and Gy of the visible light image in the horizontal and vertical directions, respectively, to obtain the gradient amplitude G of the visible light image; S3-2, perform non-maximum suppression on the calculated gradient amplitude image to refine the contour edges of objects in the key area and remove false edges. Compare the gradient direction of each pixel (i', j'), retain the local maximum value, and suppress other non-maximum points; Then the double threshold method is used to detect and connect edges, where two thresholds T are set h and T l They are used to detect the strong edge of the object contour in the key area and the weak edge of the object contour in the key area. For the gradient amplitude G, if G>T h , then the pixel point (i', j') is judged as a strong edge point and is directly retained. l ≤G≤T h , then the pixel point (i', j') is determined to be a weak edge point, and it is necessary to further determine whether it is connected to a strong edge point. If it is connected, it is retained, otherwise it is suppressed. If G<T l , then directly suppress the (i',j').

6. The UAV adaptive microwave radar navigation system according to claim 5 is characterized by: The process of extracting the contour of the object in the key area by the edge detection unit further includes: in the process of edge detection, reducing edge detection error to improve the accuracy of edge detection, which performs the following operations: S3-3. For the overlapping area between adjacent image blocks during edge detection, its width is set to N pixels. Then, for each overlapping area, a linearly increasing weight function w(x, y) is defined to ensure the smoothness of the edge fusion of the object contour in the key area: x, y are the position index of each pixel (i', j') in the overlapping area. end ,y end ) and (x start ,y start ) represent the end and start position indexes of the overlapping area respectively; S3-4. For each pixel point (i', j') in the overlapping area, the corresponding weight value is calculated according to its position index x, y, and the edge values ​​of adjacent image blocks are fused using the weighted average method: In the formula, E merged (x, y) is the merged edge value, E i (x, y) is the edge value of the i-th image block at position index (x, y), w i (x, y) is the weight value of the i-th image block at the position index (x, y), M is the number of image blocks involved in the fusion, where for non-overlapping areas, the original edge value is used directly without weighting processing; S3-5. Compare the weighted fused edge image with the original edge image to check whether there is obvious error or distortion. If it is found that the edge error in a certain area is still large, adjust the weight distribution according to the actual situation.

7. The UAV adaptive microwave radar navigation system according to claim 5 is characterized by: The feature extraction module also includes: The image texture feature extraction unit is used to analyze the texture information of the visible light image and extract texture features to distinguish different terrains and objects. The process is as follows: S3-6, select a local window W of appropriate size, W = n*n pixels, slide it on the edge image processed by step S3-5, and for each area covered by window W, set its window center coordinates to (x w ,y w ), then the pixel coordinate range in the window is: At this time, calculate the mean gradient amplitude in the window If there are N non-overlapping windows on the image, the window center coordinates are (x wi ,y wi ), i = 1, 2, ..., N, calculate the average value of the gradient amplitude in all windows Then, the standard deviation of the mean gradient amplitude in all windows is calculated As a measure of texture density of texture information characterizing edge images, the difference between the texture change degrees of different local areas of edge images is obtained.

8. The UAV adaptive microwave radar navigation system according to claim 1, characterized in that: The target recognition module comprises: The target recognition unit is used to identify and classify the target objects in the image using the image recognition algorithm. The process is as follows: First, prepare the visible light images after distortion correction, edge detection, and texture feature extraction preprocessing, collect image data with magnitudes related to the target object, and annotate them to construct a training data set; Secondly, using the convolutional neural network (CNN) architecture in deep learning, the training data set is input into the CNN model for training. During this training process, the model parameters are continuously adjusted to minimize the loss function so that the model can learn the characteristics of the target object. Again, after the training is completed, the visible light image to be identified is input into the trained CNN model, the target objects in the image are identified and classified, and the category information of the target objects is output.

9. The UAV adaptive microwave radar navigation system according to claim 1, characterized in that: The process of the image fusion module generating a comprehensive image representing accurate environmental information is as follows: First, the visible light image I after the previous processing is v (x, y) and radar data R(m, n) are normalized to make the data range of the two uniform; Secondly, establish the state model x k+1 =F k *x k +B k *u k +w k With the observation model z vk =H vk *x k +v vk 、z rk =H rk *x k +v rk , used for Kalman filtering, where the state model describes the motion law of the target object, and the observation model represents the position information of the target object obtained through the image; Again, in the Kalman filter iteration, in each frame image, the Kalman filter algorithm is used to predict the state of the target object at the current moment based on the state at the previous moment, and then the predicted state is updated in combination with the observation information of the target object in the current frame image to obtain the accurate position of the target object in the current frame; and the covariance estimate is updated to determine the fusion weight a; From this, according to formula I fused (x,y)=a*I vnorm (x,y)+(1-a)*R norm (x, y) The normalized visible light image and radar data are fused to generate a comprehensive image, where a is determined according to the Kalman filter result; Finally, the composite image is post-processed using Formula I final (x,y)=I fused (x,y)*(max final -min final )+min final Restore the pixel values ​​to the original range to obtain I final (x, y) serves as an environmental information image that provides more accurate environmental information to the navigation information generation module.

10. The UAV adaptive microwave radar navigation system according to claim 9, characterized in that: The navigation information generating module comprises: A position estimation unit estimates the current position of the UAV based on the features in the environmental information image and the position information of the target object; The heading calculation unit calculates the heading of the drone according to the position of the target object and the current position of the drone, and provides guidance for the flight direction of the drone. The process is as follows: first, the target object is identified from the environmental information image, and its position coordinates xtarget, ytarget in the image coordinate system are obtained; secondly, the coordinates are converted into the position in the actual geographic coordinate system, and the conversion is performed through the mapping relationship between the pre-calibrated image and the geographic coordinates, and then the current position xdrone, ydrone of the drone is obtained based on the position estimation unit; thirdly, according to the position coordinates of the target object and the drone, the heading angle θ of the drone relative to the target object is calculated using trigonometric functions, θ=arctan2(ytarget-ydrone, xtarget-xdrone), where arctan2 is a four-quadrant inverse tangent function; finally, the calculated heading angle is adjusted to meet the flight control requirements of the drone, and the final heading angle is output as part of the navigation information to provide guidance for the flight direction of the drone; Obstacle avoidance information generation unit: By analyzing the obstacle information in the environmental information image, it generates obstacle avoidance instructions, controls the UAV to avoid obstacles, and ensures flight safety.

Images