Unmanned aerial vehicle landing positioning method based on visual recognition

Abstract

The application discloses a kind of unmanned aerial vehicle landing positioning methods based on visual identification, it is related to positioning technical field.The application includes the multi-level image partition structure of the current aerial picture based on being constructed and including center candidate area and peripheral auxiliary area group, the regional feature of each image partition is extracted, and candidate landing point feature set is formed;For each landing area unit in candidate landing point feature set, image feature evolution sequence in continuous time frame is collected, and is consistent with the trend of backtracking generation consistency score parameter;For each landing area unit, the image feature of different angle of peripheral collaborative flight equipment is collected, and is aligned with the structure of current device area feature, and constructs space mutual proof matrix;Based on trend consistency score parameter and space mutual proof matrix, joint weighted fusion operation is executed, and landing guide vector is generated.The application realizes high-precision, anti-interference, low dependency and strong environmental adaptability of unmanned aerial vehicle autonomous landing visual positioning process.

Description

Technical Field

[0001] This invention relates to the field of positioning technology, and in particular to a method for locating a drone landing based on visual recognition. Background Technology

[0002] With the widespread application of drone technology, vision-based autonomous landing systems have become a key functional module in drone flight control systems. Especially in environments with weakened or obstructed GPS signals, visual recognition has become a crucial means of ensuring safe drone landings. Currently, most visual positioning systems rely on preset landmarks or markers for target identification and location matching. By extracting image features and combining them with coordinate transformation models, the location of landmarks is determined, assisting the drone in autonomously adjusting its attitude for precise landing. However, with the increasing prevalence of multi-drone collaborative operations and complex flight scenarios, traditional visual recognition methods relying on single perspectives or static landmarks are gradually revealing limitations in robustness, real-time performance, and positioning accuracy. Particularly in situations where the drone's perspective changes drastically, or where the landing area experiences multipath interference or obstruction, traditional image recognition results are easily disturbed, leading to increased positioning errors or even landing failure.

[0003] CN107066981A discloses a visual hierarchical landmark positioning and recognition system for autonomous landing of small unmanned aerial vehicles (UAVs). This system deploys visual hierarchical landmarks within a pre-defined landing area, acquires images using an onboard camera, and analyzes the landmark information using a visual recognizer to obtain the UAV's yaw information relative to the landmarks, thereby correcting the landing position. This method, to some extent, compensates for the instability of GPS signals and can effectively assist the landing behavior of small UAVs. However, it heavily relies on the pre-determined accuracy of physical landmarks. If landmarks are obstructed, distorted, or deployed incorrectly, the recognition algorithm is prone to failure, making it difficult to adapt to real-time extraction and dynamic judgment of landing targets in open or unstructured environments. Furthermore, this method primarily relies on single-device image information and cannot achieve multi-view verification and spatial structure collaboration, resulting in weak tolerance for image misidentification and limiting the system's application scalability in complex environments.

[0004] CN112486207A discloses a vision-based autonomous landing method for unmanned aerial vehicles (UAVs). By combining binocular camera calibration and landmark recognition with a distance-difference positioning strategy, it improves the accuracy and adaptability of landmark recognition. Furthermore, through image processing algorithms such as edge contour detection and connected component analysis, it achieves a hierarchical recognition path from coarse positioning to fine matching. Although this method can progressively improve landing accuracy over multiple distances, its positioning mechanism still relies on the clarity and structural integrity of explicit markers (such as feature-coded landmarks). When the landing area cannot be pre-marked, or when the markers are damaged or blurred, the accuracy and stability of the landing process still face challenges. In addition, this method does not introduce a cross-device collaborative verification mechanism in feature judgment. When image features are affected by ambient lighting and angle changes, it cannot effectively eliminate misjudgments or noise responses, easily leading to false landing guidance. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention is proposed.

[0006] Therefore, the problem to be solved by this invention is how to achieve dual verification of the trend consistency and spatial structure of candidate landing areas, so as to effectively improve the robustness and multi-angle adaptation capability of image feature recognition.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a visual recognition-based method for UAV landing positioning, comprising: constructing a multi-level image partitioning structure consisting of a central candidate area and a surrounding auxiliary area based on the current aerial image; extracting regional features of each image partition to form a candidate landing point feature set; for each landing area unit in the candidate landing point feature set, collecting the image feature evolution sequence in consecutive time frames and generating a trend consistency score parameter by reverse backtracking; for each landing area unit, calling the image features from different angles collected by surrounding cooperative flight equipment and aligning them structurally with the regional features in the current equipment to construct a spatial mutual verification matrix; and performing a joint weighted fusion operation based on the trend consistency score parameter and the spatial mutual verification matrix to generate a landing guidance vector.

[0009] As a preferred embodiment of the visual recognition-based UAV landing positioning method of the present invention, the construction of the multi-level image partitioning structure includes: acquiring the current aerial image; determining the center projection coordinate point as the image partitioning anchor point based on the camera intrinsic parameters and the UAV position and attitude calculation results; and setting n concentric ring region structures R1 to R2 with equal or adaptive spacing, starting from the center projection coordinate point. n As a primary boundary framework for multi-level image partitioning; each concentric annular region R kThe image is divided into sector-shaped units according to the adaptive azimuth angle interval, and then summarized to form an image partition set.

[0010] As a preferred embodiment of the visual recognition-based UAV landing positioning method of the present invention, the formation of the candidate landing point feature set includes: extracting regional features of each image partition region in the image partition set, wherein the regional features include texture structure vectors and geometric morphology parameters; based on the high-dimensional feature set composed of regional features, generating a regional discreteness index according to the distribution density and directional consistency variation value of each image partition region in the current aerial image, and marking image partition regions with an index threshold as landing region units to form a candidate landing point feature set.

[0011] As a preferred embodiment of the visual recognition-based UAV landing positioning method of the present invention, the image feature evolution sequence includes: for each landing region unit in the candidate landing point feature set, setting a fixed-length image time window, constructing a time frame buffer structure corresponding to the region index, the time frame buffer structure being used to sequentially store image sub-blocks located at spatial index positions in the image frame sequence; the region index being the spatial index position of the landing region unit in the current aerial image; based on the time frame buffer structure, extracting the texture gradient change rate and structural boundary variation of each image sub-block respectively, and constructing an image feature evolution sequence in chronological order; the texture gradient change rate being obtained through the relative change of local binary mode features; the structural boundary variation being obtained based on a weighted fusion of edge map changes and HOG histogram changes; the reverse backtracking includes: setting a backtracking reference anchor point at the end of the image feature evolution sequence, using a set reverse step size, performing a step-by-step backscan in the image feature evolution sequence, extracting feature segments with gradually decreasing structural change rates, and defining the feature segments as structural backtracking reference sequences.

[0012] As a preferred embodiment of the visual recognition-based UAV landing positioning method of the present invention, the generation of trend consistency score parameters includes: comparing the image feature evolution sequence and the structural backtracking reference sequence in a time-series sliding window manner, calculating the proportion of continuous similarity intervals and the offset distance of structural inflection points, and outputting trend consistency score parameters T1.

[0013] As a preferred embodiment of the visual recognition-based UAV landing positioning method of the present invention, the structural alignment with the regional features in the current device includes: generating a regional location information set based on the layer index and image coordinates of each landing area unit in the candidate landing point feature set; receiving image segments from a cooperative device with responsive capabilities that match the coordinate positions in the regional location information set, extracting the texture structure vector and geometric morphology parameters from the image segments to form an image segment feature set; and performing a structural image alignment operation on each image segment in the image segment feature set based on the regional features of each image partition region in the image partition set, using the principles of boundary contour stretching consistency and minimum principal direction deviation.

[0014] As a preferred embodiment of the visual recognition-based UAV landing positioning method of the present invention, the construction of the spatial mutual verification matrix includes: extracting aligned structural similarity indices from the candidate landing point feature set and the different angle feature set, and summarizing them to form a spatial mutual verification matrix T2; the structural similarity indices are obtained by weighted combination scoring of the SSIM structural similarity index and the geometric morphology consistency measure.

[0015] As a preferred embodiment of the visual recognition-based UAV landing positioning method of the present invention, the step of performing joint weighted fusion operation to generate landing guidance vector includes: fusing the regional structure similarity index under the one-to-one correspondence in the trend consistency score parameter T1 and the spatial mutual verification matrix T2 according to the regional index to construct a joint regional score vector; constructing a regional linkage graph represented by a graph structure based on the spatial distribution relationship of the candidate landing point feature set, calculating the dynamic score coordination degree between each landing region unit and its adjacent landing region units in the regional linkage graph, and forming a dynamic coordination matrix; performing fusion at the positions of region index i and adjacent region j in the joint regional score vector and the dynamic coordination matrix to generate a regional fusion score matrix; obtaining a set of region index sequences through a maximum density path search strategy according to the regional fusion score matrix, and constructing a continuous regional center path based on the image coordinate centroid of each image partition region; calculating the direction vector between the center points of adjacent image partition regions according to the continuous regional center path, forming a direction vector sequence, and generating an initial path direction vector using a weighted average strategy, wherein the weight is the joint score value of each image partition region; and projecting the initial path direction vector onto the center position of the current aerial image to form a landing guidance vector.

[0016] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, they implement the steps of the visual recognition-based UAV landing positioning method as described in the first aspect of the present invention.

[0017] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, they implement the steps of the visual recognition-based UAV landing and positioning method described in the first aspect of the present invention.

[0018] The beneficial effects of this invention are as follows: By constructing a multi-level image partitioning structure and combining image feature extraction and region division of the central candidate area and surrounding auxiliary areas, this invention effectively improves the spatial accuracy and structural analysis capability of landing point selection. Based on a time-series analysis mechanism, it collects image feature evolution sequences and combines them with trend consistency scoring parameters from reverse backtracking, achieving highly robust identification of dynamically stable regions and improving positioning stability. Furthermore, this invention significantly enhances the redundancy verification capability and spatial accuracy of visual information through a multi-angle image collaborative verification mechanism, based on feature alignment from multiple devices and multiple perspectives and the construction of a spatial mutual verification matrix. Finally, a highly reliable landing guidance vector is generated through a joint weighted fusion strategy, balancing local feature stability and global path rationality. Overall, this invention achieves a high-precision, anti-interference, low-dependency, and highly environmentally adaptable autonomous landing visual positioning process for UAVs, particularly suitable for complex terrain, GPS signal-limited, or multi-UAV collaborative scenarios. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a vision-based drone landing and positioning method. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] As mentioned in the background section, most current visual positioning systems rely on preset landmarks or markers for target recognition and location matching. By extracting image features and combining them with coordinate transformation models, the location of landmarks is determined, assisting UAVs in autonomously adjusting their attitude for precise landing. However, with the increasing number of multi-drone collaborative operations and complex flight scenarios, traditional visual recognition methods relying on single perspectives or static landmarks are gradually revealing limitations in terms of robustness, real-time performance, and positioning accuracy. Especially in situations where the UAV's perspective changes drastically, or where there is multi-path interference or occlusion in the landing area, traditional image recognition results are easily disturbed, leading to increased positioning errors or even landing failure.

[0025] Figure 1 This is a flowchart of a visual recognition-based UAV landing and positioning method according to an embodiment of the present invention. Figure 1 As shown, the visual recognition-based UAV landing localization method includes:

[0026] S1: Based on the current aerial footage, construct a multi-level image partitioning structure consisting of a central candidate area and a surrounding auxiliary area, extract the regional features of each image partition, and form a set of candidate landing point features.

[0027] In this embodiment of the invention, the construction of the multi-level image partitioning structure includes:

[0028] 1. Acquire the current aerial image, and determine the center projection coordinate point based on the camera intrinsic parameters and the drone's position and attitude calculation results, which will serve as the image partition anchor point.

[0029] The current aerial images are acquired by the drone's downward-facing camera during flight, which continuously captures images at timestamp intervals. The current aerial images are a two-dimensional image data matrix in RGB or YUV format.

[0030] Camera intrinsic parameters include fixed parameters such as focal length, principal point coordinates, and distortion coefficients, which are used to describe the optical imaging characteristics of the camera itself.

[0031] The position and attitude parameters of the UAV include three-dimensional orientation angle information measured by a three-axis accelerometer and gyroscope, and geographical location information output by satellite positioning.

[0032] After dynamic fusion using algorithms such as Kalman filtering, the camera's projection transformation matrix can be obtained. By back-projecting the target point in the three-dimensional ground reference coordinate system to the two-dimensional image coordinate system using the projection transformation matrix, the center projection coordinate point on the image plane can be calculated.

[0033] 2. Taking the central projection coordinate point as the starting point, set up n concentric ring region structures R1 to R2 with equal or adaptive spacing. n This serves as the primary boundary framework for multi-level image partitioning; where n is the set number of vertical region levels in the image, dynamically adjusted according to image resolution and height.

[0034] Specifically, each annular region maintains a constant pixel spacing or an adaptive dynamic spacing with its adjacent inner and outer regions. A constant pixel spacing means there is a fixed radius difference between regions, while the adaptive dynamic spacing is adjusted based on changes in the drone's flight altitude, image resolution, and scene complexity. For example, if the flight altitude is low or the target area is dense, a smaller radius step size is set to improve region resolution; if the flight altitude is high or the image is sparse overall, a larger step size can be used to improve partition coverage efficiency.

[0035] The number of vertical region layers *n* is a positive integer representing the number of concentric ring regions generated from the center region to the image edge. The value of *n* is dynamically set based on the maximum side length of the current aerial image, the image sampling interval, and scene detail information, generally adjusted between 3 and 7. Compared to traditional methods that only divide image regions into rectangles or grids, this invention constructs a ring-shaped region boundary framework, forming a spatially progressive structure between the image center and edges. This facilitates highlighting the central candidate region while maintaining the image distribution structure of the edge auxiliary regions, thus enhancing the spatial relevance of candidate point extraction.

[0036] 3. Divide each concentric annular region R k The image is divided into sector-shaped units according to the adaptive azimuth angle interval, and then summarized to form an image partition set, thus completing the multi-level image partition structure of the current aerial image.

[0037] The specific operations include: using the central projection coordinate point as the pole, dividing the entire 360° circular annular region into several fan-shaped image sub-regions with equal or adaptive angles according to the angle interval θ. Equal angle division refers to fixing the angle interval of each fan-shaped region, for example, dividing a region into 60° segments to obtain 6 fan-shaped units; adaptive angle division dynamically adjusts according to the changes in texture density or orientation gradient in the scene, using smaller angle intervals in areas with drastic texture changes to improve extraction accuracy.

[0038] By dividing the image into two dimensions (n ​​layers in the radial direction and m segments in the angular direction), the original image can be divided into n×m image partition units. Each image partition unit has attributes such as spatial number, image coordinate boundaries, and relative center angle.

[0039] The resulting image partition set possesses spatial structure hierarchical features and directional distribution balance, which not only enhances the directional sensitivity of region selection but also gives different weights to the central and surrounding regions in the candidate point evaluation.

[0040] In this embodiment of the invention, the formation of the candidate landing point feature set includes:

[0041] 4. Calculate the regional features of each image partition in the image partition set. The regional features include texture structure vector (based on gradient co-occurrence texture histogram) and geometric morphology parameters (including indicators such as principal direction, boundary closure, and regional uniformity).

[0042] Specifically, the method for extracting texture structure vectors is based on gradient co-occurrence texture histogram (GCH), which involves calculating the gradient of the grayscale image within the region, constructing a grayscale gradient co-occurrence matrix, and generating a texture statistical vector of a fixed dimension based on this matrix.

[0043] The texture statistics vector includes gradient direction distribution density, local texture roughness index, and texture isotropic measure.

[0044] The geometric morphology parameters include three main dimensions: principal direction refers to the principal component direction of all boundary gradient directions within the region, used to characterize the structural consistency within the region; boundary closure indicates the continuity and integrity of the region's contour edge, which is evaluated by the degree of closure of the boundary pixel chain; and region uniformity refers to the statistical consistency of the texture distribution within the region, using indicators such as standard deviation to measure the dispersion of pixel values.

[0045] By combining texture structure vectors and geometric morphology parameters, a high-dimensional region feature vector is constructed, which has the ability to describe the spatial structure of the region and the texture of the image, and provides a feature basis for subsequent candidate evaluation.

[0046] 5. Based on the high-dimensional feature set composed of regional features, calculate the distribution density and directional consistency variation value of each image partition region in the current aerial image, generate the regional dispersion index, and mark the image partition regions with higher than the index threshold as landing area units to form a candidate landing point feature set.

[0047] It should be noted that the distribution density is defined as the sum of the similarity between the feature vector of a region and itself within a specified range (such as a 5×5 adjacent region) around each image partition. The higher the value, the more consistent the corresponding region is with its adjacent regions in terms of structure.

[0048] The orientation consistency variation value refers to the standard deviation of the main orientation of the current image partition region from the main orientation of the adjacent regions. The smaller the value, the more stable the structure and the more regular the orientation of the region.

[0049] The regional dispersion index is defined as a weighted combination score of directional consistency variation value and distribution density, which evaluates the structural stability and spatial aggregation ability of each image region.

[0050] S2: For each landing region unit in the candidate landing point feature set, collect the image feature evolution sequence in consecutive time frames, and generate trend consistency score parameters by reverse backtracking.

[0051] S2.1: For each landing area unit in the candidate landing point feature set, set a fixed-length image time window (e.g., set to 5 to 15 frames, which can be dynamically adjusted according to the frame rate and flight speed of the UAV, and this invention does not limit it), and construct a time frame buffer structure corresponding to the area index.

[0052] The time frame buffer structure is used to sequentially store image sub-blocks located at spatial index positions in the image frame sequence.

[0053] The region index is the spatial index position of the landing area unit in the current aerial image.

[0054] It should be noted that conventional methods, which only process texture analysis of static images, cannot capture the dynamic impact of environmental disturbances or structural changes on candidate regions. Setting an image time window allows for the establishment of an image change trend model on the time axis, which is a crucial foundation for improving the accuracy of landing point stability judgment. This invention, through a time frame buffer structure, ensures the spatial consistency of regions in the time dimension. Even with slight camera shake or flight drift, perspective transformation correction can be used to align region index positions across frames. Compared to conventional frame-by-frame independent processing, the time frame buffer structure features fixed spatial indices and synchronous evolution of image sub-blocks.

[0055] S2.2: To ensure that the region correspondence remains consistent on the image timeline, based on the time frame buffer structure, the texture gradient change rate and structural boundary variation of each image sub-block are extracted, and an image feature evolution sequence is constructed in chronological order.

[0056] The texture gradient change rate is obtained through the relative change of the local binary pattern (LBP) features. Specifically, after grayscale processing of each image sub-block, a 3×3 (for example only) neighborhood is extracted with each pixel as the center. The grayscale values ​​of the center pixel and the neighboring pixels are compared to construct an 8-bit binary code to form a local binary pattern map. The LBP map is normalized and histogram statistics are performed to obtain the texture vector of each frame's image sub-block. The texture change rate between two frames is defined as the chi-square distance of the LBP histogram divided by the time difference, which represents the rate of change of the texture structure.

[0057] By arranging the LBP change rate between consecutive frames in a time series, a texture change gradient sequence is obtained.

[0058] The structural boundary variation is obtained through a weighted fusion of edge map changes and HOG histogram changes. The specific calculation process is as follows: the edge map extracts image boundary information using Canny or Sobel operators, representing the boundary distribution as a binary image; the edge map variation is calculated by performing a pixel-level XOR operation between the current frame's edge map and the previous frame's edge map, and then counting the number of pixels with boundary changes. The HOG descriptor is used to characterize the local gradient direction distribution features. The image is divided into small blocks, and a gradient direction histogram is calculated in each block, normalized to form an HOG vector. The Euclidean distance between the HOG vectors of the current frame and the previous frame represents the structural directional change.

[0059] The structural boundary variation is a weighted average of the edge map variation rate and the HOG variation rate. The weights can be set according to the complexity of the region's structure, such as increasing the weight of edge variation in regions with complex textures.

[0060] For each image sub-block in the image time window, the texture gradient change rate and structural boundary variation of the corresponding frame are extracted according to the time order of the frame number. The texture gradient change rate and structural boundary variation of each frame are combined into a set of two-dimensional feature vectors, and the sequence is serialized to obtain the image feature evolution sequence.

[0061] In addition, to eliminate high-frequency noise interference and scale inconsistency, it is necessary to normalize each vector value in the sequence (such as Z-score normalization) and perform sliding window averaging to enhance the recognizability of trend features.

[0062] S2.3: Set a backtracking reference anchor point at the end of the image feature evolution sequence, and perform stepwise backscanning in the image feature evolution sequence by setting a reverse step size to extract feature segments with gradually decreasing structural change rate, and define the feature segments as structural backtracking reference sequences.

[0063] The specific operations include: setting the backtracking reference anchor point to the last frame image sub-block in the time frame buffer structure, representing the image state at the latest time point; setting the backscan step size, for example, backtracking once every 1 or 2 frames; reviewing the image feature evolution sequence frame by frame according to the step size; and performing a trend-decreasing search of the structural change rate.

[0064] During the retracement process, if the texture change rate and structural boundary variation are detected to decrease synchronously or fall below the stability threshold for multiple consecutive frames (this can be achieved by statistically analyzing the evolution sequences of image features from a large number of known stable regions and calculating the standard deviation of the structural change rate fluctuation range), then this time window is marked as a stable evolution segment. The image structure within this time segment exhibits a convergence or steady-state trend, which is called the structural backtracking reference sequence.

[0065] Unlike conventional overall mean statistical methods, the backtracking segment exhibits a decreasing trend over time, which can effectively eliminate the impact of short-term fluctuations or random disturbances on the score and improve the reliability of trend judgment.

[0066] S2.4: Generating trend consistency scoring parameters T1 includes: comparing the structural trend of the image feature evolution sequence and the structural backtracking reference sequence in a time-series sliding window manner, calculating the proportion of continuous similarity intervals and the offset distance of structural inflection points, and outputting trend consistency scoring parameters T1.

[0067] The specific operation is as follows: a sliding window of a set length is slid sequentially in the original evolution sequence, and the data in each sliding window and the corresponding sliding window in the structural backtracking segment are compared for feature similarity.

[0068] The similarity calculation consists of two parts: first, the proportion of continuous similarity intervals, representing the percentage of windows in the sliding window whose similarity scores are higher than the similarity threshold (which can be set according to actual operation); and second, the structural inflection point offset distance, defined as the maximum offset in the time dimension between the structural abrupt change point in the current evolution trend and the corresponding inflection point in the retrospective sequence. Specifically, first-order derivative sequences are constructed for the texture gradient change rate and the structural boundary variation, respectively; and a joint index of their change rates is constructed based on the square root of the sum of their squares, reflecting the overall change rate of texture and boundary structure in each time frame, forming a change rate sequence. In the change rate sequence, if an element has a change rate greater than that of the elements before and after it, it is judged as a candidate point for structural abrupt change, and compared with a set abrupt change threshold. Only structural abrupt change candidate points that exceed the set abrupt change threshold (which can be set as the mean of the joint index of change rates plus one standard deviation) are retained to form a set of structural inflection points.

[0069] These two indicators are weighted and calculated to generate a trend consistency score parameter, which serves as a quantitative basis for the structural stability and evolutionary consistency of the current landing area unit over time. A higher trend consistency score parameter value indicates that the current area has good historical structural stability and is suitable for landing point determination in dynamic environments.

[0070] S3: For each landing area unit, call the image features from different angles collected by surrounding cooperative flight equipment, and perform structural alignment with the area features in the current equipment to construct a spatial mutual verification matrix.

[0071] S3.1: Generate a set of regional location information based on the layer index and image coordinates of each landing area unit in the candidate landing point feature set.

[0072] The layer index is a combination of the ring level number and the angle partition number of each sector unit when constructing a multi-level image partition structure, which can identify the topological relationship of the region's position in the multi-level structure.

[0073] Image coordinates include the pixel coordinates of the top-left corner of the landing region unit in the current image frame, as well as the pixel width and height of the region, denoted as a quadruple, forming a complete two-dimensional region bounding box. Based on the above structure, a region location information set is constructed, which is an ordered structure containing the layer indices of all candidate regions and image boundary parameters.

[0074] To ensure the uniqueness and coordination of regional features, the regional location information set also needs to be bound to a timestamp and a spatial identifier: the timestamp identifies the frame time in which the region was extracted in this aerial footage, while the spatial identifier is generated by the UAV's location information (latitude and longitude, flight altitude, heading angle) and layer index, and is encoded through a hash or hash function so that subsequent coordinating devices can achieve accurate queries in time and space.

[0075] S3.2: Receive image sub-blocks at different angles from image segments fed back by cooperative devices with responsive capabilities, whose coordinate positions match those in the regional location information set, extract texture structure vectors and geometric morphological parameters from the image sub-blocks at different angles, and form a set of features at different angles.

[0076] The coordinating devices include similar or dissimilar flight terminals deployed at different locations in the airspace, at different flight altitudes or attitude angles. Upon receiving a location request, these devices extract image sub-blocks from local caches or real-time image frame sequences that match the image coordinates in the regional location information set.

[0077] Because of the differences in viewing angles between flight equipment, image sub-blocks are observations of the current landing area unit from different angles, and are called different-angle image sub-blocks. The extraction of image sub-blocks not only considers pixel coordinates, but also needs to combine the camera extrinsic matrix. The target area is projected onto the current viewing angle through homography matrix or 3D projective transformation to complete coordinate matching and affine region calibration, ensuring that the different-angle image is spatially aligned with the original area patch.

[0078] S3.3: For each image sub-block with different angles in the set of different angle features, the principle of consistent boundary contour stretching and minimum deviation of main direction is adopted to perform structural image alignment operation based on the regional features of each image partition region in the set of image partitions.

[0079] Ideally, the alignment principle in this step is based on two core criteria: consistency of boundary profile stretching and minimum deviation in the principal direction.

[0080] Among them, boundary contour stretching consistency is used to measure whether the contour shape of the image sub-blocks at different angles after projective transformation is close to the geometric stretching degree of the candidate region contour in the current device image. It is calculated as the affine transformation strain energy between the corresponding point sets of the contour (e.g., average point distance error or stretching ratio).

[0081] The principle of minimizing principal orientation deviation is used to measure the angle difference between the principal orientation angles (the angles corresponding to the maximum values ​​of the orientation gradients) of two image sub-block structures, and cosine or Euclidean distance is used as the scoring standard.

[0082] The two criteria are weighted and combined to form a structural alignment similarity value, where a larger structural alignment similarity value indicates more consistent structural alignment. Within each image partition in the image partition set, the structural alignment similarity value with each image sub-block at a different angle is calculated. The region with the highest alignment similarity is selected as the current matching region, and the corresponding structural features are subjected to synchronous normalization processing to achieve structural alignment and information integration of heterogeneous images.

[0083] Compared with traditional single feature matching or key point alignment methods, the structural alignment method of this invention has dual guarantees of geometric integrity and directional stability, and adapts to the target region verification needs under complex perspectives.

[0084] S3.4: Extract aligned structural similarity indices from the candidate landing point feature set and the different angle feature set, and summarize them to form a spatial mutual verification matrix T2; wherein, the structural similarity index is obtained by weighted combination scoring of SSIM structural similarity index and geometric morphological consistency measure.

[0085] Specifically, the structural similarity index uses a weighted combination of two dimensions: the first is the structural similarity index SSIM, which measures the consistency of brightness, contrast and structural distribution; the second is the geometrical similarity index, which includes geometrical descriptive indicators such as boundary length, principal axis direction, curvature and area.

[0086] All structural similarity values ​​are organized into a matrix structure and arranged by region index to form a spatial mutual verification matrix. Each element represents the degree of structural consistency of the current region under multi-view cross-verification.

[0087] S4: Based on the trend consistency score parameter and spatial mutual verification matrix, perform a joint weighted fusion operation to generate a landing guidance vector.

[0088] S4.1: Merge the regional structure similarity indices under the one-to-one correspondence in the trend consistency scoring parameter T1 and the spatial mutual verification matrix T2 according to the regional index to construct a joint regional scoring vector.

[0089] The joint regional score vector is defined as the weighted calculation result of the trend consistency score parameter and the structural similarity index in the spatial mutual verification matrix.

[0090] S4.2: Based on the spatial distribution relationship of the candidate landing point feature set, construct a regional linkage graph represented by a graph structure, calculate the dynamic score coordination degree of each landing area unit with its adjacent landing area units in the regional linkage graph, and form a dynamic coordination matrix.

[0091] The specific method is to construct an undirected graph G = (V, E) based on the geometric centroid coordinates of each image partition region in the aerial image, where V is the set of nodes composed of all candidate landing area units, and E represents the adjacency relationship between nodes.

[0092] The criterion for establishing adjacency is: if the centroid distance d between two regions is greater than or equal to the distance between their centroids... ij If the region spacing is less than the threshold (e.g., 1 / 10 of the image width), an edge is created. After the undirected graph structure is built, the rating coordination degree between each pair of adjacent region units in the undirected graph is calculated based on the joint region rating vector, generating a dynamic coordination matrix. The calculation formula is as follows:

[0093]

[0094] Among them, S i and S j σ represents the joint region score for image partitions with region indices i and j, respectively, and σ is the distance attenuation coefficient.

[0095] S4.3: In the joint regional score vector and dynamic coordination matrix, perform fusion at the positions of regional index i and adjacent regional j to generate a regional fusion score matrix.

[0096] The regional fusion scoring matrix is ​​the result of combining the joint regional scoring vector and the dynamic coordination matrix. The construction method is as follows: for all regional index pairs (i,j) with adjacency in the undirected graph structure, extract the joint score value corresponding to region i from the joint regional scoring vector; extract the score value of the regional index pair (i,j) from the dynamic coordination matrix; and then weight and fuse the above two score values ​​through a weight factor to construct the elements of the regional fusion scoring matrix.

[0097] S4.4: Based on the region fusion scoring matrix, a set of region index sequences are obtained through the maximum density path search strategy, and continuous region center paths are constructed by sequentially connecting them based on the image coordinate centroid of each image partition region.

[0098] The path search process can employ heuristic greedy path expansion or dynamic programming methods. Starting from the node with the highest score in the graph, the path expands sequentially to the adjacent node with the best score until a path sequence with a length that reaches a preset threshold or covers the central region of the image is formed.

[0099] The objective function for path search is defined as maximizing the total path fusion score. The resulting region sequence not only has high local credibility but also ensures structural continuity and directional consistency between regions, which can be used to represent the proposed landing guidance route.

[0100] S4.5: Based on the center path of the continuous region, calculate the direction vector between the center points of adjacent image partition regions to form a direction vector sequence, and use a weighted average strategy to generate the initial path direction vector, where the weight is the joint score value of each image partition region.

[0101] S4.6: Project the initial path direction vector onto the center of the current aerial image to form a landing guidance vector.

[0102] Specifically, the initial path direction vector generated by the weighted average is spatially projected. Starting from the center coordinates of the current aerial image, with the direction being the initial path direction vector and the length being a set unit vector scale, a landing guidance vector trajectory is drawn. This trajectory guides the UAV to fly towards the optimal regional fusion path and ultimately perform a stable landing. This guidance method offers higher spatial continuity, robustness, and synergy compared to traditional methods based on single-point or regular image feature extraction.

[0103] This embodiment also provides a computer device applicable to the UAV landing positioning method based on vision recognition, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the UAV landing positioning method based on vision recognition as proposed in the above embodiment.

[0104] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0105] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the visual recognition-based drone landing positioning method proposed in the above embodiments.

[0106] In summary, this invention effectively improves the spatial accuracy and structural analysis capability of landing point selection by constructing a multi-level image partitioning structure and combining image feature extraction and region division of the central candidate area and surrounding auxiliary areas. Based on a time-series analysis mechanism, it collects image feature evolution sequences and combines them with trend consistency scoring parameters from reverse backtracking, achieving highly robust identification of dynamically stable regions and improving positioning stability. Furthermore, this invention significantly enhances the redundancy verification capability and spatial accuracy of visual information through a multi-device, multi-view feature alignment and spatial mutual verification matrix construction. Finally, a highly reliable landing guidance vector is generated through a joint weighted fusion strategy, balancing local feature stability and global path rationality. Overall, this invention achieves a high-precision, anti-interference, low-dependency, and highly environmentally adaptable autonomous landing visual positioning process for UAVs, particularly suitable for complex terrain, GPS signal-limited, or multi-UAV collaborative scenarios.

[0107] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for UAV landing and positioning based on visual recognition, characterized in that: include: Based on the current aerial footage, a multi-level image partitioning structure consisting of a central candidate area and a surrounding auxiliary area is constructed. The regional features of each image partition are extracted to form a set of candidate landing point features. For each landing region unit in the candidate landing point feature set, the image feature evolution sequence in consecutive time frames is collected, and the trend consistency score parameter is generated by reverse backtracking. For each landing area unit, the features of images from different angles collected by surrounding cooperative flight equipment are called up and structurally aligned with the regional features in the current equipment to construct a spatial mutual verification matrix. Based on the trend consistency score parameter and the spatial mutual verification matrix, a joint weighted fusion operation is performed to generate a landing guidance vector. The image feature evolution sequence includes: for each landing region unit in the candidate landing point feature set, setting a fixed-length image time window, constructing a time frame buffer structure corresponding to the region index, the time frame buffer structure being used to sequentially store image sub-blocks located at spatial index positions in the image frame sequence; the region index is the spatial index position of the landing region unit in the current aerial image; based on the time frame buffer structure, extracting the texture gradient change rate and structural boundary variation of each image sub-block respectively, and constructing an image feature evolution sequence in chronological order; the texture gradient change rate is obtained through the relative change of local binary mode features; the structural boundary variation is obtained based on a weighted fusion of edge map changes and HOG histogram changes; The reverse backtracking includes: setting a backtracking reference anchor point at the end of the image feature evolution sequence, using a set reverse step size, performing a step-by-step backscan in the image feature evolution sequence, extracting feature segments whose structural change rate gradually slows down, and defining the feature segments as structural backtracking reference sequences; The generation of trend consistency scoring parameters includes: comparing the image feature evolution sequence and the structural backtracking reference sequence using a time-series sliding window method to calculate the proportion of continuous similarity intervals and the offset distance of structural inflection points, and outputting trend consistency scoring parameters. ; Extract aligned structural similarity indices from the candidate landing point feature set and the feature set from different angles, and summarize them to form a spatial mutual verification matrix. ; The process of performing a joint weighted fusion operation to generate a landing guidance vector includes: incorporating trend consistency score parameters... Mutual verification matrix of space The regional structural similarity index under the one-to-one correspondence relationship is fused according to the regional index to construct a joint regional scoring vector. Based on the spatial distribution relationship of the candidate landing point feature set, a regional linkage graph represented by a graph structure is constructed. The dynamic scoring coordination degree of each landing region unit with its adjacent landing region units in the regional linkage graph is calculated to form a dynamic coordination matrix. In the joint regional scoring vector and the dynamic coordination matrix, fusion is performed at the positions of region index i and adjacent region j to generate a regional fusion scoring matrix. According to the regional fusion scoring matrix, a set of regional index sequences is obtained through the maximum density path search strategy, and continuous regional center paths are constructed by connecting them sequentially based on the image coordinate centroid of each image partition region. According to the continuous regional center paths, the direction vector between the center points of adjacent image partition regions is calculated to form a direction vector sequence, and an initial path direction vector is generated using a weighted average strategy, where the weight is the joint scoring value of each image partition region. The initial path direction vector is projected onto the center position of the current aerial image to form a landing guidance vector.

2. The UAV landing positioning method based on visual recognition as described in claim 1, characterized in that: The construction of the multi-level image partitioning structure includes: Acquire the current aerial image, and determine the center projection coordinate point based on the camera intrinsic parameters and the drone's position and attitude calculation results, which will serve as the image partition anchor point; Starting from the central projection coordinate point, set A concentric ring region structure with equal or adaptive spacing. to , serving as the primary boundary framework for multi-level image partitioning; Each concentric ring-shaped region The image is divided into sector-shaped units according to the adaptive azimuth angle interval, and then summarized to form an image partition set.

3. The UAV landing positioning method based on visual recognition as described in claim 2, characterized in that: The formation of the candidate landing point feature set includes: Extract the regional features of each image partition region in the image partition set, the regional features including texture structure vectors and geometric morphological parameters; Based on a high-dimensional feature set composed of regional features, a regional discreteness index is generated according to the distribution density and directional consistency variation value of each image partition in the current aerial image. Image partitions with values ​​higher than the index threshold are marked as landing area units, forming a candidate landing point feature set.

4. The UAV landing positioning method based on visual recognition as described in claim 1, characterized in that: The structural alignment with regional features in the current device includes: Based on the layer index and image coordinates of each landing area unit in the candidate landing point feature set, a set of regional location information is generated. Receive image sub-blocks at different angles from image segments fed back by cooperative devices with responsive capabilities, which match the coordinate positions in the regional location information set; extract the texture structure vectors and geometric morphology parameters from the image sub-blocks at different angles to form a set of features at different angles. For each image sub-block with different angles in the set of different angle features, the principle of consistent boundary contour stretching and minimum deviation of main direction is adopted, and a structural image alignment operation is performed based on the regional features of each image partition region in the set of image partitions.

5. The UAV landing positioning method based on visual recognition as described in claim 4, characterized in that: The construction of the spatial mutual verification matrix includes: The structural similarity index is obtained by weighted combination scoring of the SSIM structural similarity index and geometrical consistency measure.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the visual recognition-based UAV landing and positioning method according to any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the visual recognition-based UAV landing and positioning method according to any one of claims 1 to 5.

Images