Unmanned aerial vehicle parking system

Abstract

The invention discloses an unmanned aerial vehicle parking system, and relates to the technical field of unmanned aerial vehicles, the unmanned aerial vehicle parking system comprises an unmanned aerial vehicle body, and a sensing module, a processing module and a transmission module which are arranged on the unmanned aerial vehicle body, the sensing module is a binocular vision acquisition module, the processing module is an embedded computing power processing module, and the transmission module is connected with the processing module. Parking area images are acquired through the binocular vision acquisition module, depth estimation, BEV visual angle conversion, semantic segmentation, quantitative parameter calculation and other processing are executed in combination with the embedded computing power processing module, microdefects and liquid pollutants on the surface of the parking area can be accurately and quantitatively evaluated, the safety state of the parking area is clearly judged, and the safety of the parking area is improved. Therefore, a comprehensive and accurate surface state basis is provided for parking of the unmanned aerial vehicle, potential safety hazards such as sideslip and clamping stagnation caused by an unknown surface state are effectively avoided, the stability and safety of the unmanned aerial vehicle in the parking process are guaranteed, and the actual application requirement for high-precision safe parking of the unmanned aerial vehicle is met.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, specifically to a UAV docking system. Background Technology

[0002] With the continuous development of drone technology, its application in various fields such as logistics delivery, power line inspection, and environmental monitoring is becoming increasingly widespread. Autonomous drone docking, as a crucial link in the operational process, directly impacts overall operational efficiency and equipment safety in terms of safety and reliability. In practical applications, the surface condition of the docking area is one of the core factors determining docking safety. Therefore, accurate assessment of the surface condition of the docking area is an important prerequisite for achieving safe and autonomous drone docking.

[0003] A method and system for dynamic landing point adjustment of UAVs based on environmental perception, disclosed in CN120235065A, is proposed. This scheme acquires environmental perception data and flight status parameters, constructs a three-dimensional terrain mesh model by using a fusion scanning method of lidar and vision, analyzes the terrain undulation distribution and obstacle density distribution, generates a set of safe landing areas, and then adjusts the landing point position and landing strategy through a multi-path optimization model. This existing technology mainly achieves dynamic adjustment of landing points through macro-terrain features and obstacle distribution, which improves the adaptability of UAV landing in complex terrain to a certain extent.

[0004] However, the above-mentioned solutions can only achieve qualitative or semi-quantitative analysis of macroscopic terrain undulations and obstacle density, and cannot accurately and quantitatively assess the microscopic defects and liquid contaminants on the surface of the docking area that affect docking stability. Therefore, they cannot provide a comprehensive and accurate basis for the surface condition of UAVs and cannot meet the actual needs of high-precision and safe UAV docking. Therefore, developing a UAV docking system that can accurately and quantitatively assess the microscopic defects and liquid contaminants on the surface of the docking area has become an urgent technical problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a drone docking system to solve the problem that the prior art cannot accurately and quantitatively assess the microscopic defects and liquid contaminants on the surface of the docking area that affect the docking stability, thus failing to provide a comprehensive and accurate basis for the surface condition of drones for docking.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a drone docking system, comprising a drone body and a sensing module, a processing module, and a transmission module disposed on the drone body, wherein the sensing module is a binocular vision acquisition module and the processing module is an embedded computing power processing module;

[0007] It also includes an auxiliary calibration module. The binocular vision acquisition module is used to acquire left and right eye images of the docking area and transmit them to the embedded computing power processing module. The auxiliary calibration module is used to provide a calibration benchmark. The embedded computing power processing module is configured to perform the following operations: preprocess the left and right eye images; perform depth estimation based on the preprocessed images to generate a depth map; combine the depth map with the calibration benchmark provided by the auxiliary calibration module to perform BEV viewpoint conversion; perform semantic segmentation on the converted BEV image to extract surface defects and liquid contaminant features; perform quantitative parameter calculation based on the extracted features and depth map; and perform docking safety determination based on the quantitative parameter calculation results. The transmission module is used to transmit the safety determination results to the flight control system of the UAV itself.

[0008] Furthermore, when the embedded computing power processing module performs depth estimation, it employs an improved semi-global block matching algorithm. This algorithm introduces gradient information weights to optimize the cost calculation function and sets a reflective area mask to fill the disparity value of the reflective area with the neighborhood mean. The depth estimation process also includes dynamic disparity range adaptation processing, and the adaptation relationship satisfies the formula: ,in This refers to the parallax range after dynamic adaptation. As the reference parallax range, This refers to the actual flight altitude of the drone. The reference height is used for system calibration; the auxiliary calibration module includes three calibration markers, and the embedded computing power processing module performs intrinsic and extrinsic parameter calibration of the binocular vision acquisition module by acquiring images of the calibration markers. The intrinsic parameters include focal length, principal point coordinates and distortion coefficients, and the extrinsic parameters include rotation matrix and translation vector.

[0009] Furthermore, when the embedded computing power processing module performs BEV perspective transformation, it uses an inverse projection algorithm to map image pixels to the XY plane of the world coordinate system to generate a BEV image. The mapping relationship of the inverse projection algorithm satisfies the formula: ,in and Let X be the coordinates of the pixel in the XY plane of the world coordinate system. and These are the x and y coordinates of the pixel in the image coordinate system. and Principal point coordinates This represents the depth value corresponding to the pixel. and The focal length of the camera in the X and Y axes is given, and the pixel accuracy of the BEV image has a fixed mapping relationship with the physical size of the world coordinate system.

[0010] Furthermore, when the embedded computing power processing module performs semantic segmentation, it adopts an improved lightweight U-Net model based on MobileNetV3 as the backbone network. The model replaces the ordinary convolutional layers of the original U-Net with depthwise separable convolutional layers and introduces a boundary enhancement loss function to strengthen the extraction of boundary features of surface defects and liquid contaminants. The output of the semantic segmentation includes background category, surface defect category, pit category, and liquid contaminant category. The embedded computing power processing module is also configured to post-process the semantic segmentation results to remove isolated pixel regions with an area smaller than a set threshold.

[0011] Furthermore, the quantitative parameter calculation includes the calculation of quantitative parameters for surface defects, including cracks and pits; the quantitative parameter calculation for cracks includes crack width calculation, which is performed by selecting multiple measurement points along the crack direction in the BEV image, calculating the crack pixel width at each measurement point, converting it to the actual width, and then taking the average value; the quantitative parameter calculation for pits includes pit depth calculation, which is performed by extracting the depth value of the pit area in the depth map, performing a difference operation with the depth value of the surrounding normal area, and taking the maximum value of the difference as the pit depth. The calculation relationship satisfies the formula: ,in The depth of the pit. This is the depth value of the surrounding normal area. This represents the depth of the pit area.

[0012] Furthermore, the quantitative parameter calculation includes the calculation of quantitative parameters for liquid pollutants, which include water accumulation and oil pollution; the quantitative parameter calculation for oil pollution includes the calculation of oil pollution coverage area, which is calculated by statistically analyzing the number of pixels in the oil pollution area in the BEV image and combining the mapping relationship between pixel accuracy and physical size to convert it into the actual oil pollution coverage area; the quantitative parameter calculation for water accumulation includes the calculation of water accumulation thickness, which is calculated by combining the depth value of the water accumulation area and the camera optical axis angle through a refraction model.

[0013] Furthermore, when the embedded computing power processing module performs docking safety determination, it presets surface defect safety thresholds and liquid contaminant safety thresholds, and compares the calculated quantitative parameters with the corresponding safety thresholds respectively; when all quantitative parameters are less than or equal to the corresponding safety thresholds, the docking area is determined to be safe to dock; when any quantitative parameter is greater than the corresponding safety threshold, the docking area is determined to be unsafe to dock; the transmission module transmits the safety determination result and quantitative parameters to the flight control system together. When it is determined to be unsafe to dock, the embedded computing power processing module is also configured to call the pre-stored backup docking point coordinates and transmit them synchronously to the flight control system through the transmission module.

[0014] Furthermore, the system also includes a supplementary lighting module, which is electrically connected to the embedded computing power processing module. The embedded computing power processing module is configured to detect the illumination intensity of the image during image preprocessing, and when the illumination intensity is lower than a set value, control the supplementary lighting module to turn on and adjust the brightness of the supplementary lighting. The image preprocessing includes grayscale processing, Gaussian filtering processing, and histogram equalization processing. The grayscale processing uses a weighted average method, and the Gaussian filtering processing uses a fixed-size convolution kernel for noise removal.

[0015] Furthermore, when the embedded computing power processing module performs intrinsic and extrinsic parameter calibration, it adopts the Zhang Zhengyou calibration method. By acquiring images of calibration markers at different angles, calibration equations are constructed to solve for the intrinsic and extrinsic parameters. The calibration markers are regular black and white patterns. The embedded computing power processing module determines the image coordinates of the markers by identifying their contour features, and then establishes a mapping relationship between the image coordinate system and the world coordinate system. After calibration, the embedded computing power processing module stores the calibration parameters in the local storage unit for use in depth estimation and BEV viewpoint conversion.

[0016] Furthermore, the transmission module is a Wi-Fi 6 transmission module, and the embedded computing power processing module, the binocular vision acquisition module, the transmission module, and the auxiliary calibration module all use industrial-grade communication protocols for data interaction; the embedded computing power processing module is also configured to verify the image data and calculation results during transmission, and trigger a retransmission mechanism when a data transmission error is detected; after receiving the safety judgment result, the flight control system of the UAV body performs docking path adjustment or backup docking point switching operations according to the judgment result.

[0017] Compared with existing technologies, this drone docking system has the following advantages:

[0018] I. This invention acquires images of the docking area through a binocular vision acquisition module, and performs depth estimation, BEV perspective conversion, semantic segmentation, and quantitative parameter calculation by an embedded computing power processing module. This enables precise quantitative assessment of microscopic defects and liquid contaminants on the surface of the docking area, clearly determining the safety status of the docking area. This provides a comprehensive and accurate basis for the surface condition of the UAV, effectively avoiding safety hazards such as sideslip and jamming caused by unknown surface conditions, ensuring the stability and safety of the UAV docking process, and meeting the practical application requirements for high-precision and safe UAV docking.

[0019] Second, through the synergistic effect of adaptive adjustment of the supplementary lighting module, lightweight algorithm optimization, and efficient data transmission mechanism, this invention can improve the system's adaptability and real-time data processing in complex environments such as different lighting and weather conditions. At the same time, it can quickly provide feedback on safety judgment results and backup docking point information, thereby helping the UAV to adjust its docking strategy in a timely manner, improving overall operational efficiency, expanding the UAV's applicability in multiple scenarios, and reducing the risk of equipment damage and operational delays caused by docking failures.

[0020] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0022] Figure 1 This is a diagram of the docking system module architecture of the present invention;

[0023] Figure 2 This is a flowchart illustrating the safety assessment process for the docking area in this invention. Detailed Implementation

[0024] To further illustrate the technical means and effects adopted by the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on the specific implementation methods, structures, features and effects of the present invention.

[0025] A drone docking system includes a drone body, a binocular vision acquisition module, an embedded computing power processing module, a transmission module, an auxiliary calibration module, and a supplementary lighting module. The binocular vision acquisition module, transmission module, auxiliary calibration module, and supplementary lighting module are all fixedly connected to the embedded computing power processing module. The embedded computing power processing module interacts with the drone's flight control system through the transmission module. Data transmission between all modules uses industrial-grade communication protocols to ensure the stability and reliability of data transmission.

[0026] The drone body is a conventional multi-rotor drone on the market. Its flight control system has basic functions such as path planning, attitude adjustment and receiving external commands. This system does not improve the basic flight functions of the drone body, but only realizes docking-related decisions and controls through data interaction.

[0027] The binocular vision acquisition module uses an industrial-grade binocular camera with a resolution of 1920×1080 pixels, a frame rate of 30fps, and a lens focal length of 8mm. It is installed in the center below the drone body with the lens facing directly downwards to ensure complete acquisition of image information of the potential docking area directly below and around the drone. The core function of this module is to acquire the left and right eye images of the docking area, providing raw image data for subsequent depth calculation and surface condition analysis.

[0028] The embedded computing power processing module uses a high-performance processor based on the ARM architecture with a main frequency of 1.8GHz. It is equipped with 4GB of RAM and 32GB of local storage, which has powerful parallel computing and data storage capabilities. It can efficiently execute a series of complex algorithms such as image preprocessing, depth estimation, BEV viewpoint conversion, semantic segmentation, quantitative parameter calculation and safety determination. The local storage unit is used to store calibration parameters, semantic segmentation model weights, spare docking point coordinates and various intermediate calculation results.

[0029] The transmission module adopts a Wi-Fi 6 transmission module, which supports a maximum transmission rate of 1.2Gbps and a communication distance of up to 100 meters. It can quickly and stably transmit safety judgment results, quantitative parameters and backup docking point coordinates to the UAV flight control system. At the same time, it can receive auxiliary information such as the actual flight altitude and flight attitude of the UAV from the flight control system, meeting the real-time transmission requirements.

[0030] The auxiliary calibration module includes three calibration markers, which are black and white squares with a side length of 20cm. The black and white areas are arranged alternately, and each marker has a circular positioning hole in the center for easy identification by the embedded computing module. The three calibration markers are arranged in an equilateral triangle on a flat surface, with a center-to-center distance of 2 meters between adjacent markers. This arrangement provides a stable and reliable calibration benchmark for the intrinsic and extrinsic parameters of the binocular vision acquisition module, ensuring the accuracy of subsequent depth estimation and coordinate transformation.

[0031] The supplementary lighting module uses a ring-shaped LED supplementary lighting array, consisting of 12 high-brightness LED beads, which is installed around the lens of the binocular vision acquisition module and electrically connected to the embedded computing power processing module. Under the precise control of the embedded computing power processing module, the core function of the supplementary lighting module is to provide an auxiliary light source for image acquisition in low-light environments, ensuring the quality of image acquisition.

[0032] Before the system is put into use, the intrinsic and extrinsic parameters of the binocular vision acquisition module need to be calibrated using the auxiliary calibration module. The calibration process adopts the Zhang Zhengyou calibration method, and the specific steps are as follows:

[0033] Step 1: Select a flat, unobstructed ground as the calibration site. Fix the three calibration points in an equilateral triangle arrangement, ensuring that the points are placed stably and without tilting, and that the center-to-center distance between adjacent points does not exceed ±5cm.

[0034] Step 2: Hover the drone 5 meters directly above the center of the equilateral triangle formed by the three calibration markers. This height is the reference height Hcalib for system calibration. Adjust the drone's attitude to keep it horizontal, with the binocular vision acquisition module's lens pointing vertically towards the ground.

[0035] Step 3: Control the binocular vision acquisition module to acquire images of the calibration markers from different angles, including 0°, 30°, 60°, 90°, 120°, and 150°. Acquire 2 images from each angle, for a total of 12 calibration images. Ensure that the three calibration markers in each image are complete, clear, unobstructed, and free from reflections.

[0036] Step 4: The embedded computing power processing module receives the collected calibration images, preprocesses each image, extracts the contour features of the calibration markers through the edge detection algorithm, identifies the black and white square contours and the central positioning hole, and then determines the coordinates of each calibration marker in the image coordinate system.

[0037] Step 5: Based on the image coordinates of the marker points and the known actual spatial coordinates of the marker points (pre-calculated according to the arrangement), construct calibration equations and solve for the intrinsic and extrinsic parameters of the binocular vision acquisition module using the least squares method. The intrinsic parameters include the camera's focal length along the X-axis. Focal length in the Y-axis direction Principal point coordinates and And distortion coefficients k1, k2, p1, p2, and extrinsic parameters including rotation matrix R and translation vector T.

[0038] Step Six: After calibration, the embedded computing power processing module stores the obtained intrinsic and extrinsic parameters in the local storage unit. These calibration parameters can be directly called during subsequent depth estimation and BEV viewpoint conversion, eliminating the need for repeated calibration. If the installation position of the binocular vision acquisition module changes or the drone collides during system use, the above calibration process must be repeated to ensure parameter accuracy.

[0039] When a drone needs to perform a docking operation, the system initiates a docking area assessment process, the specific steps of which are as follows:

[0040] (I) Image Acquisition and Preprocessing

[0041] The drone flight control system, based on the preset docking area coordinates, controls the drone to fly directly above the docking area at a specific altitude. The flight control system provides real-time feedback to the embedded computing module, and the altitude can be adjusted within the range of 1 to 10 meters according to actual operational needs.

[0042] The embedded computing power processing module controls the binocular vision acquisition module to start, simultaneously acquiring left and right eye images of the docking area. During the acquisition process, the embedded computing power processing module detects the light intensity of the image in real time. The light intensity is obtained by calculating the average grayscale value of the image. The light intensity threshold is set to 120. When the average grayscale value of the image is detected to be lower than 120, it is determined that the light is insufficient. The embedded computing power processing module controls the supplementary light module to start and adjusts the supplementary light brightness according to the difference between the average grayscale value and the threshold. The larger the difference, the higher the supplementary light brightness. The supplementary light brightness is divided into 5 levels to ensure that the image acquisition quality meets the requirements of subsequent processing.

[0043] After receiving the left and right eye images, the embedded computing power processing module performs preprocessing on the images, including grayscale conversion, Gaussian filtering, and histogram equalization.

[0044] The grayscale conversion process uses a weighted average method with weighting coefficients set to R=0.299, G=0.587, and B=0.114. The color image is converted to a grayscale image using the formula Gray=0.299R+0.587G+0.114B. These weighting coefficients conform to the sensitivity characteristics of the human eye to different colors and can preserve the key features of the image.

[0045] The Gaussian filtering process uses a 3×3 convolution kernel with a standard deviation of 0.8. The convolution operation smooths the grayscale image, removes Gaussian noise, and reduces the interference of noise on subsequent depth estimation and semantic segmentation.

[0046] Histogram equalization processing employs an adaptive histogram equalization algorithm, with a contrast threshold set to 2.0. By adjusting the grayscale distribution of the image, the contrast of the image is enhanced, making the surface defects and liquid contaminant features of the docking area more prominent, which facilitates subsequent feature extraction.

[0047] (ii) Depth estimation

[0048] The embedded computing power processing module performs depth estimation on the preprocessed left and right eye images, and generates a depth map using an improved semi-global block matching algorithm. The specific implementation process is as follows:

[0049] First, epipolar correction is performed on the left and right eye images. Based on the calibrated intrinsic and extrinsic parameters, the left and right eye images are mapped to the same epipolar direction to ensure that corresponding pixels are on the same horizontal line, thus reducing the difficulty of matching.

[0050] An improved semi-global block matching algorithm is used to calculate disparity values. This algorithm introduces gradient information weights to optimize the cost calculation function based on the traditional semi-global block matching algorithm. Specifically, the horizontal and vertical gradients of the image are calculated using the Sobel operator, and the gradient information weights are determined according to the magnitude of the gradient values. The larger the gradient value, the higher the weight. The weight coefficient is set to 0.3. The cost calculation function is a weighted sum of the initial cost function and the gradient information weight term, i.e., Cost = Cost0 + 0.3 × GradWeight. This optimization method can improve the matching accuracy of texture-rich regions.

[0051] For reflective areas in the image, the range of reflective areas is determined by detecting pixel areas with gray values ​​higher than 250, and a reflective area mask is generated. The disparity values ​​within the reflective areas are filled with the 3×3 neighborhood mean to avoid disparity calculation errors caused by reflection and improve the integrity of the depth map.

[0052] Perform dynamic parallax range adaptation processing based on the actual flight altitude of the drone. Adjust the parallax range so that the adaptation relationship satisfies the formula. Among them, the reference parallax range Set to 0-128 pixels, the reference height during system calibration. The actual flight altitude of the drone is 5 meters. Provided in real time by the flight control system, for example, the dynamically adapted parallax range when the drone's actual flight altitude is 10 meters. =128×(10 / 5)=256 pixels. Through this dynamic adaptation process, accurate parallax values ​​can be obtained at different flight altitudes.

[0053] Based on the mapping relationship between disparity values ​​and depth values, the calculated disparity values ​​are converted into depth values ​​to generate a depth map. The depth value of each pixel in the depth map is in meters, and the depth measurement accuracy can reach ±0.01 meters. The depth map can accurately reflect the three-dimensional spatial information of the docking area.

[0054] (III) BEV Perspective Shift

[0055] The embedded computing power processing module uses an inverse projection algorithm to map the pixels in the depth map to the XY plane of the world coordinate system, generating a BEV image. The specific implementation process is as follows:

[0056] Define the mapping relationship of the inverse projection algorithm, which satisfies the formula. ,in, and Let X be the coordinates of the pixel in the XY plane of the world coordinate system. and These are the x and y coordinates of the pixel in the image coordinate system. and Principal point coordinates This represents the depth value corresponding to the pixel. and This represents the focal length of the camera in the X and Y axes.

[0057] A fixed mapping relationship is established between the pixel precision of the BEV image and the physical size in the world coordinate system. In this embodiment, one pixel of the BEV image corresponds to a physical size of 0.01 meters × 0.01 meters in the world coordinate system. This mapping relationship ensures that the actual physical size can be accurately calculated from the pixel information of the BEV image.

[0058] Determine the range for generating the BEV image. Centered on the projection point directly below the drone on the XY plane of the world coordinate system, generate a 5m x 5m BEV image. This range can completely cover the potential docking area of ​​the drone, ensuring no omissions.

[0059] Based on the above mapping relationship and generation range, all pixels in the depth map are mapped one by one to the XY plane of the world coordinate system to generate a BEV image. The BEV image can intuitively show the planar distribution of the parking area from a bird's-eye view, which provides convenience for subsequent semantic segmentation and quantitative parameter calculation.

[0060] (iv) Semantic segmentation

[0061] The embedded computing power processing module uses an improved lightweight U-Net model based on MobileNetV3 as the backbone network to perform semantic segmentation on BEV images. The specific implementation process is as follows:

[0062] The output categories of semantic segmentation are clearly defined, including background category, surface defect category, pit category and liquid contaminant category. Among them, the surface defect category mainly refers to crack defects, and the liquid contaminant category includes water accumulation and oil stains.

[0063] An improved lightweight U-Net model was constructed, which uses MobileNetV3-Small as the backbone network and sets the depth multiplier to 0.75. The ordinary convolutional layers of the original U-Net model were replaced with depthwise separable convolutional layers. The depthwise separable convolutional layers split the 3×3 ordinary convolution into 3×3 depthwise convolutions and 1×1 pointwise convolutions. While ensuring segmentation accuracy, the number of model parameters and computational cost were significantly reduced, the model's running speed was improved, and the real-time requirements of the embedded computing power processing module were met.

[0064] During model training, a boundary enhancement loss function is introduced. Based on the cross-entropy loss function, this function assigns higher weights (weight coefficient is set to 1.5) to the boundary regions of surface defects and liquid contaminants, thereby enhancing the model's ability to extract boundary features and making the boundaries of the segmentation results clearer and more accurate.

[0065] After normalizing the BEV image (normalizing pixel values ​​to the range of 0-1), it is input into the improved lightweight U-Net model. The model outputs semantic segmentation results, which are semantic label maps of the same size as the BEV image, with each pixel corresponding to a semantic category.

[0066] Post-processing is performed on the semantic segmentation results. The area threshold of isolated pixel regions is set to 5 pixels. The isolated pixel regions in the semantic label map are identified by the connected component analysis algorithm. Isolated pixel regions with an area of ​​less than 5 pixels are removed and classified into adjacent main semantic categories. This post-processing step can effectively remove the missegmentation points generated during the semantic segmentation process and improve the accuracy of the semantic segmentation results.

[0067] (v) Calculation of quantitative parameters

[0068] The embedded computing power processing module calculates quantitative parameters for surface defects and liquid contaminants based on semantic segmentation results and depth maps. The specific implementation process is as follows:

[0069] Calculation of quantitative parameters for surface defects

[0070] Surface defects include cracks and pits. For cracks, the crack width is calculated as follows: First, the crack area is identified in the semantic label map of the BEV image. A measurement point is selected every 0.05 meters along the crack direction. At each measurement point, a straight line perpendicular to the crack direction is drawn. The number of pixels between the intersection of this line and the boundary of the crack area is counted. The actual crack width at the measurement point is calculated based on the pixel precision of the BEV image (0.01 meters / pixel). At least 5 measurement points are selected, and the average of the actual widths of all measurement points is calculated as the final width of the crack.

[0071] For pits, pit depth calculation is performed: First, the pit region is identified in the semantic label map, and then the depth values ​​of all pixels in the pit region are extracted from the depth map. Then, using the boundary of the pit area as a reference, extend outward by 0.1 meters to determine the normal area surrounding the pit, and extract the depth values ​​of all pixels within this normal area. ,calculate The average value, according to the formula Calculate the pit depth, i.e., select... With each The maximum value of the difference is taken as the final depth of the pit, and this calculation method can accurately reflect the maximum degree of pit depression.

[0072] Calculation of quantitative parameters for liquid pollutants

[0073] Liquid contaminants include water and oil. For oil contamination, the oil contamination coverage area is calculated as follows: oil contamination areas are identified in the semantic tag map, the number of pixels in the area is counted, and the actual oil contamination coverage area is calculated using the formula S = number of pixels × 0.01 × 0.01, based on the pixel accuracy of the BEV image (0.01 meters / pixel).

[0074] For water accumulation, the following steps are performed: First, identify the water accumulation area in the semantic tag map and extract the depth values ​​of all pixels within the water accumulation area in the depth map. Then, select a dry, normal area around the water accumulation area, extract the depth values ​​of this normal area, and calculate the average value. Obtain the pitch and roll angles of the UAV (provided by the UAV's IMU sensor, which is 0° by default, i.e., the camera optical axis is perpendicular to the ground). Based on the refractive index of water (1.33), the depth difference is corrected through a refraction model. This correction process can eliminate the influence of water accumulation refraction on depth measurement and ensure the accuracy of water accumulation thickness calculation.

[0075] (vi) Determination of docking safety

[0076] The embedded computing power processing module presets safety thresholds for surface defects and liquid contaminants. The safety thresholds are set based on the docking requirements of conventional small drones, and the specific values ​​are as follows: the safety threshold for crack width is 0.02 meters, the safety threshold for pit depth is 0.03 meters, the safety threshold for oil stain coverage area is 0.1 square meters, and the safety threshold for water accumulation thickness is 0.05 meters. Users can adjust the above safety thresholds through the system interface according to the actual drone model and docking scenario requirements.

[0077] The embedded computing power processing module compares the calculated quantitative parameters with the corresponding safety thresholds: when the crack width is ≤0.02 meters, the pit depth is ≤0.03 meters, the oil stain coverage area is ≤0.1 square meters and the water accumulation thickness is ≤0.05 meters, the parking area is determined to be safe for parking; when any quantitative parameter is greater than the corresponding safety threshold, the parking area is determined to be unsafe for parking.

[0078] If the location is determined to be unsafe for docking, the embedded computing power processing module calls the coordinates of the backup docking point pre-stored in the local storage unit. The backup docking point coordinates are planned by the user in advance and input into the system. The system stores at least 3 backup docking point coordinates. The embedded computing power processing module selects the nearest backup docking point coordinates as the target coordinates based on the distance between the drone's current position and each backup docking point.

[0079] The embedded computing power processing module transmits the safety assessment results, various quantitative parameters (if the drone is safe to dock) or the coordinates of the backup docking point (if the drone is not safe to dock) to the flight control system of the drone body through the transmission module (Wi-Fi 6 transmission module). During the data transmission, the image data and calculation results are verified by CRC32 check. When a data transmission error is detected, a retransmission mechanism is triggered, which allows a maximum of 3 retransmissions. If all 3 retransmissions fail, the system issues an alarm signal to prompt the user to check the communication link.

[0080] After receiving the above data, the UAV's flight control system performs the following operations: If it receives a determination that the UAV can be safely docked, the flight control system adjusts the UAV's attitude according to the originally planned docking path and controls the UAV to land smoothly in the target docking area; if it receives a determination that the UAV cannot be safely docked and the coordinates of the alternative docking point, the flight control system replans the flight path and controls the UAV to go to the alternative docking point. After arriving directly above the alternative docking point, the system restarts the above docking area assessment process to assess the safety of the alternative docking point until a safe docking area is found. If all alternative docking points are determined to be unsafe to dock, the flight control system controls the UAV to return to the departure point or the designated safe area.

[0081] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention are within the scope of the present invention.

Claims

1. A drone docking system, comprising a drone body and a sensing module, a processing module, and a transmission module disposed on the drone body, characterized in that, The perception module is a binocular vision acquisition module, and the processing module is an embedded computing power processing module; It also includes an auxiliary calibration module. The binocular vision acquisition module is used to acquire left and right eye images of the docking area and transmit them to the embedded computing power processing module. The auxiliary calibration module is used to provide a calibration benchmark. The embedded computing power processing module is configured to perform the following operations: preprocess the left and right eye images; perform depth estimation based on the preprocessed images to generate a depth map; combine the depth map with the calibration benchmark provided by the auxiliary calibration module to perform BEV viewpoint conversion; perform semantic segmentation on the converted BEV image to extract surface defects and liquid contaminant features; perform quantitative parameter calculation based on the extracted features and depth map; and perform docking safety determination based on the quantitative parameter calculation results. The transmission module is used to transmit the safety determination results to the flight control system of the UAV itself.

2. The unmanned aerial vehicle (UAV) docking system according to claim 1, characterized in that, When the embedded computing power processing module performs depth estimation, it adopts an improved semi-global block matching algorithm. This algorithm introduces gradient information weights to optimize the cost calculation function and sets a reflective area mask to fill the disparity value of the reflective area with the neighborhood mean. The depth estimation process also includes dynamic disparity range adaptation processing, and the adaptation relationship satisfies the formula: ,in This refers to the parallax range after dynamic adaptation. As the reference parallax range, This refers to the actual flight altitude of the drone. The reference height is used for system calibration; the auxiliary calibration module includes three calibration markers, and the embedded computing power processing module performs intrinsic and extrinsic parameter calibration of the binocular vision acquisition module by acquiring images of the calibration markers. The intrinsic parameters include focal length, principal point coordinates and distortion coefficients, and the extrinsic parameters include rotation matrix and translation vector.

3. The unmanned aerial vehicle (UAV) docking system according to claim 1, characterized in that, When the embedded computing power processing module performs BEV perspective transformation, it uses an inverse projection algorithm to map image pixels to the XY plane of the world coordinate system to generate a BEV image. The mapping relationship of the inverse projection algorithm satisfies the formula: ,in and Let X be the coordinates of the pixel in the XY plane of the world coordinate system. and These are the x and y coordinates of the pixel in the image coordinate system. and Principal point coordinates This represents the depth value corresponding to the pixel. and The focal length of the camera in the X and Y axes is given, and the pixel accuracy of the BEV image has a fixed mapping relationship with the physical size of the world coordinate system.

4. The unmanned aerial vehicle (UAV) docking system according to claim 1, characterized in that, When the embedded computing power processing module performs semantic segmentation, it adopts an improved lightweight U-Net model based on MobileNetV3 as the backbone network. The model replaces the ordinary convolutional layers of the original U-Net with depth-separable convolutional layers and introduces a boundary enhancement loss function to strengthen the extraction of boundary features of surface defects and liquid contaminants. The output of the semantic segmentation includes background category, surface defect category, pit category and liquid contaminant category. The embedded computing power processing module is also configured to post-process the semantic segmentation results and remove isolated pixel regions with an area smaller than a set threshold.

5. The unmanned aerial vehicle (UAV) docking system according to claim 1, characterized in that, The quantitative parameter calculation includes the quantitative parameter calculation of surface defects, which include cracks and pits; The quantitative parameter calculation for cracks includes crack width calculation, which is performed by selecting multiple measurement points along the crack direction in the BEV image, calculating the crack pixel width at each measurement point, converting it to the actual width, and then taking the average value. The quantitative parameter calculation for potholes includes pothole depth calculation, which is performed by extracting the depth value of the pothole area from the depth map, performing a difference operation with the depth value of the surrounding normal area, and taking the maximum difference as the pothole depth. The calculation relationship satisfies the formula: ,in The depth of the pit. This is the depth value of the surrounding normal area. This represents the depth of the pit area.

6. The unmanned aerial vehicle (UAV) docking system according to claim 1, characterized in that, The quantitative parameter calculation includes the quantitative parameter calculation of liquid pollutants, which includes water and oil; the quantitative parameter calculation of oil includes the calculation of oil coverage area, which is calculated by counting the number of pixels in the oil area in the BEV image and combining the mapping relationship between pixel accuracy and physical size to convert it into the actual oil coverage area. The calculation of quantitative parameters for water accumulation includes the calculation of water accumulation thickness. The water accumulation thickness calculation combines the depth value of the water accumulation area and the camera optical axis angle, and is performed through a refraction model.

7. The unmanned aerial vehicle (UAV) docking system according to claim 1, characterized in that, When the embedded computing power processing module performs docking safety determination, it presets surface defect safety thresholds and liquid contaminant safety thresholds, and compares the calculated quantitative parameters with the corresponding safety thresholds respectively; when all quantitative parameters are less than or equal to the corresponding safety thresholds, the docking area is determined to be safe to dock; when any quantitative parameter is greater than the corresponding safety threshold, the docking area is determined to be unsafe to dock. The transmission module transmits the safety assessment result and quantitative parameters to the flight control system. When it is determined that the landing is not safe, the embedded computing power processing module is also configured to call the pre-stored backup landing point coordinates and transmit them to the flight control system synchronously through the transmission module.

8. The unmanned aerial vehicle (UAV) docking system according to claim 1, characterized in that, The system also includes a supplementary lighting module, which is electrically connected to an embedded computing power processing module. The embedded computing power processing module is configured to detect the illumination intensity of the image during image preprocessing. When the illumination intensity is lower than a set value, it controls the supplementary lighting module to turn on and adjust the brightness of the supplementary lighting. The image preprocessing includes grayscale processing, Gaussian filtering processing, and histogram equalization processing. The grayscale processing uses a weighted average method, and the Gaussian filtering processing uses a fixed-size convolution kernel for noise removal.

9. A drone docking system according to claim 2, characterized in that, When the embedded computing power processing module performs intrinsic and extrinsic parameter calibration, it adopts the Zhang Zhengyou calibration method. By acquiring images of calibration markers at different angles, calibration equations are constructed to solve for the intrinsic and extrinsic parameters. The calibration markers are regular black and white patterns. The embedded computing power processing module determines the image coordinates of the markers by identifying their contour features, and then establishes a mapping relationship between the image coordinate system and the world coordinate system. After calibration, the embedded computing power processing module stores the calibration parameters in the local storage unit for use in depth estimation and BEV viewpoint conversion.

10. The unmanned aerial vehicle (UAV) docking system according to claim 1, characterized in that, The transmission module is a Wi-Fi 6 transmission module. The embedded computing power processing module, the binocular vision acquisition module, the transmission module, and the auxiliary calibration module all use industrial-grade communication protocols for data interaction. The embedded computing power processing module is also configured to verify the image data and calculation results during transmission. When a data transmission error is detected, a retransmission mechanism is triggered. After receiving the safety judgment result, the flight control system of the UAV body performs docking path adjustment or backup docking point switching operations according to the judgment result.

Images