A method and system for real-time stitching and target annotation of UAV aerial images

Abstract

This application discloses a real-time stitching and target annotation method and system for UAV aerial images, relating to the field of UAV aerial image processing technology. The method includes: calling the initial pose parameters and global spatial reference of a continuous aerial image sequence; performing coordinate system alignment processing on the initial pose parameters and global spatial reference to generate a coordinate alignment sequence; performing contour stability verification on the continuous aerial image sequence based on the coordinate alignment sequence, separating deformation interference regions and fixed structure regions, and generating deformation state parameters; determining spatial deviation based on the deformation state parameters; generating region association weights based on the spatial deviation; determining pose compensation parameters based on the region association weights; performing global mapping correction on the initial pose parameters based on the pose compensation parameters to generate a corrected stitching sequence; and outputting annotation position information based on the corrected stitching sequence and target region identifiers. This application improves the accuracy of aerial image stitching and the spatial positioning reliability of target annotation in fire rescue scenarios.

Description

Technical Field

[0001] This application relates to the field of UAV aerial image processing technology, and in particular to a method and system for real-time stitching and target annotation of UAV aerial images. Background Technology

[0002] During fire rescue operations in super high-rise buildings, the continuous frame sequences captured by drones during low-altitude aerial photography missions need to be stitched together and labeled in real time to provide spatial reference for fire command. Conventional aerial photography stitching methods typically treat all edge features in the continuous frame sequence as fixed structural references, directly participating in the mapping and calculation of the global coordinate system.

[0003] Conventional methods lack a mechanism to separate dynamic boundaries from static references in the contour extraction and pose calculation stages, making it impossible to distinguish between contour migration caused by thermal deformation and the actual pose drift of the UAV at the data level. The stitching process directly accumulates the displacement data of the deformation interference area into the homography transformation matrix, causing the global coordinate system to shift in tandem with local thermal expansion.

[0004] Therefore, conventional aerial stitching methods cannot prevent the deformation boundary from exerting a pulling effect on the global mapping when the building facade undergoes continuous thermal deformation, resulting in cross-regional misalignment of the target annotation position during thermal expansion. Summary of the Invention

[0005] In view of the aforementioned problems, this application is hereby filed.

[0006] Therefore, this application provides a method and system for real-time stitching and target annotation of UAV aerial images, which can solve the problems mentioned in the background art.

[0007] To solve the above-mentioned technical problems, this application provides the following technical solution: In a first aspect, this application provides a method for real-time stitching and target labeling of UAV aerial images, including: receiving aerial stitching control instructions and target area identifiers, calling the initial pose parameters and global spatial reference of a continuous aerial frame sequence, performing coordinate system alignment processing on the initial pose parameters and the global spatial reference, and generating a coordinate alignment sequence; Based on the coordinate alignment sequence, the contour stability of the continuous aerial frame sequence is checked to separate the deformation interference region from the fixed structure region and generate deformation state parameters. Based on the deformation state parameters, the spatial deviation is determined, the regional association weight is generated based on the spatial deviation, and the pose compensation parameters are determined based on the regional association weight. The initial pose parameters are globally mapped and corrected according to the pose compensation parameters to generate a corrected stitching sequence. The annotation position information is then output based on the corrected stitching sequence and the target region identifier.

[0008] Preferably, the step of performing contour stability verification on the continuous aerial frame sequence based on the coordinate alignment sequence, separating the deformation interference region from the fixed structure region, and generating deformation state parameters includes: Extract the edge contour lines of adjacent frames in the continuous aerial frame sequence, determine the curvature change trend and displacement gradient distribution between the edge contour lines of adjacent frames, and generate a contour evolution matrix; By comparing the inter-frame displacement differences of each contour region in the contour evolution matrix, the contour regions in which the displacement differences are unidirectionally accumulated along the frame order direction are classified as deformation interference regions, and the contour regions in which the displacement differences are constant along the frame order direction are classified as fixed structure regions. The boundary pixels contained in the deformation interference region are traversed, and the offset distance and morphological distortion of the boundary pixels between adjacent frames are obtained. The deformation state parameters are then generated by combining these parameters.

[0009] Preferably, the step of determining the spatial deviation based on the deformation state parameters, generating region association weights based on the spatial deviation, and determining pose compensation parameters based on the region association weights includes: The deformation state parameters are mapped to the corresponding pixel grid of the continuous aerial frame sequence, and the spatial displacement data between the corresponding pixel grid and the fixed structure region is obtained to generate local spatial deviation. The attenuation coefficient field is determined based on the modulus distribution of the local spatial deviation. The attenuation coefficient field and the mapping confidence of the corresponding pixel grid are subjected to spatial superposition processing to generate region association weights. The numerical range of the associated weights of the region is uniformly processed, and the pose compensation parameters are determined by combining the rigid transformation parameters of the fixed structural region.

[0010] Preferably, determining the curvature change trend and displacement gradient distribution between the edge contour lines of adjacent frames, and generating a contour evolution matrix, includes: Extract contour pixels along the edge contour lines of the adjacent frames at a fixed step size, perform neighborhood smoothing on the contour pixels, and generate a smooth contour sequence. The contour direction of each pixel is determined according to the smooth contour sequence, the contour direction is converted into a tangent direction vector, and a corresponding normal direction vector is generated based on the tangent direction vector. The contour evolution matrix is ​​generated by combining the tangent direction vector and the normal direction vector.

[0011] Preferably, the step of determining the attenuation coefficient field based on the modulus distribution of the local spatial deviation, and performing spatial superposition processing on the mapping confidence of the attenuation coefficient field and the corresponding pixel grid to generate region association weights includes: The distribution of extreme points of the local spatial deviation in the global pixel grid is statistically analyzed to determine the deformation diffusion center and the deformation static boundary. A radial attenuation field is determined starting from the deformation diffusion center, and an attenuation coefficient is generated based on the distribution relationship between the relative distance from the global pixel grid to the deformation diffusion center and the length of the deformation static boundary. The attenuation coefficient and the mapping confidence of the corresponding pixel grid are subjected to spatial domain fusion processing to output the region association weight.

[0012] Preferably, the step of performing global mapping correction on the initial pose parameters according to the pose compensation parameters, generating a corrected stitching sequence, and outputting annotation location information according to the corrected stitching sequence and the target region identifier includes: The pose compensation parameters are decomposed into translation compensation components and rotation compensation components, and spatial transformation processing is performed on the initial pose parameters respectively to generate a corrected pose sequence. Based on the corrected pose sequence, perform grid deformation coordinate mapping processing on the continuous aerial frame sequence to generate a corrected stitched sequence; The pixel coordinates of the corrected stitched sequence are transformed to a globally unified coordinate system, and the corresponding annotation position information is determined according to the target region identifier, and the annotation position information is output.

[0013] Preferably, the step of performing neighborhood smoothing processing on the contour pixels to generate a smooth contour sequence includes: Obtain the spatial distribution neighborhood of the contour pixels, and construct a neighborhood distribution ratio matrix that matches the size of the spatial distribution neighborhood; The neighborhood distribution ratio matrix and the coordinate values ​​of the contour pixels are weighted and fused to generate smooth pixel coordinates. The smooth pixel coordinates are rearranged in order according to the fixed step size to obtain a smooth contour sequence.

[0014] Preferably, the step of performing spatial domain fusion processing on the mapping confidence of the attenuation coefficient and the corresponding pixel grid, and outputting the region association weight, includes: Construct a spatial weight distribution array that matches the size of the continuous aerial frame sequence, and fill the central region of the spatial weight distribution array with the attenuation coefficients according to the spatial distribution; The mapping confidence is combined with the spatial weight distribution matrix to perform feature superposition processing, thereby generating a local weight distribution matrix; The local weight distribution matrix is ​​subjected to neighborhood numerical smoothing based on the sliding window range, and the region-related weights are output.

[0015] Preferably, the step of performing mesh deformation coordinate mapping processing on the continuous aerial frame sequence based on the corrected pose sequence to generate a corrected stitched sequence includes: The continuous aerial frame sequence is subjected to mesh division processing based on the corrected pose sequence to generate an initial stitched mesh. The corrected pose sequence is mapped to the initial stitched mesh, and the global projection error of the initial stitched mesh is determined. Based on the global projection error distribution, perform mesh vertex displacement adjustment processing to generate a deformation compensation mesh; The pixel content of the continuous aerial frame sequence is resampled into the corresponding topology of the deformation compensation grid, and the corrected stitching sequence is output.

[0016] Secondly, this application also provides a real-time stitching and target labeling system for UAV aerial images, including: an alignment module, which receives aerial stitching control instructions and target area identifiers, calls the initial pose parameters and global spatial reference of a continuous aerial frame sequence, performs coordinate system alignment processing on the initial pose parameters and the global spatial reference, and generates a coordinate alignment sequence; The verification module performs contour stability verification on the continuous aerial frame sequence based on the coordinate alignment sequence, separates the deformation interference region from the fixed structure region, and generates deformation state parameters. The compensation module determines the spatial deviation based on the deformation state parameters, generates regional association weights based on the spatial deviation, and determines pose compensation parameters based on the regional association weights. The correction module performs global mapping correction on the initial pose parameters according to the pose compensation parameters, generates a corrected stitching sequence, and outputs annotation position information based on the corrected stitching sequence and the target region identifier.

[0017] Implementing this application will have the following beneficial effects: 1. This application divides the unidirectional accumulation region into a deformation interference region and the periodic fluctuation region into a fixed structure region by performing inter-frame displacement difference accumulation calculation and dual threshold interval classification. Based on the spatial deviation, a radial attenuation distribution field is constructed and spatial fusion processing is performed with the mapping confidence to generate the region association weight. The thermal deformation boundary is identified and stripped to avoid dynamic deformation data from being mistakenly involved in the global pose calculation. This solves the problem of global coordinate system linkage offset caused by the default fixed building boundary in conventional splicing, and achieves physical isolation of dynamic and static features.

[0018] 2. This application performs spatial transformation processing on the initial pose by decomposing the pose compensation parameters into translation and rotation components, performs mesh generation based on the corrected pose sequence and quantizes the global projection error node by node, performs mesh vertex displacement adjustment to generate a deformation compensation mesh according to the error distribution, and resamples the pixel content to the deformation compensation mesh to output the corrected stitching sequence. This avoids the spatial drift of the target annotation coordinates with the false boundary, solves the problem of cross-regional misalignment of the target annotation position during thermal expansion, and improves the aerial stitching accuracy and spatial positioning reliability of the target annotation in fire rescue scenarios. Attached Figure Description

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

[0020] Figure 1 This is an overall flowchart of a method for real-time stitching and target annotation of drone aerial images related to this application; Figure 2 This is an application environment diagram of a real-time stitching and target annotation method for UAV aerial images involved in this application; Figure 3 This is a schematic diagram of the overall structure of a real-time stitching and target annotation system for drone aerial images involved in this application; Figure 4 This is a computer device diagram of a method for real-time stitching and target annotation of drone aerial images, which is the subject of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0022] In one exemplary embodiment, such as Figure 1 As shown, a method for real-time stitching and target annotation of UAV aerial images is provided, including: S1: Receive aerial photography stitching control instructions and target area identifiers, call the initial pose parameters and global spatial reference of the continuous aerial photography frame sequence, perform coordinate system alignment processing on the initial pose parameters and global spatial reference, and generate a coordinate alignment sequence; It should be noted that the aerial stitching control commands are issued by the UAV ground control terminal or the airborne flight management system, and the commands carry the mission execution timestamp and data acquisition frequency band parameters. The target area identifier refers to the zoning code or geofence coordinate range of the building facades that require key monitoring at the fire rescue site. The initial pose parameters of the continuous aerial frame sequence are acquired in real time by the UAV's inertial measurement unit and satellite positioning module, including the aircraft's three-dimensional spatial coordinates, pitch angle, roll angle, and yaw angle. The global spatial reference refers to the pre-mapped three-dimensional point cloud data of the building's main structure or the set of coordinates of fixed control points. Coordinate system alignment processing aims to eliminate the initial deviation between the sensor-acquired coordinate system and the actual geographic coordinate system of the building.

[0023] Specifically, the processing system parses the timestamp and frequency band parameters in the aerial photography stitching control command and retrieves a continuous sequence of aerial frames that matches the target area identifier. It extracts the initial pose parameters corresponding to the continuous aerial frame sequence and projects these parameters onto the coordinate plane of the global spatial reference. It calculates the translation vector between the origin of the initial pose parameter coordinate system and the origin of the global spatial reference coordinate system, and calculates the rotation matrix between the axes of the initial pose parameter coordinate system and the axes of the global spatial reference coordinate system. The translation vector and rotation matrix are superimposed on the original pixel data of the continuous aerial frame sequence, and a spatial reference registration operation is performed to generate a coordinate-aligned sequence. The coordinate system alignment process establishes a one-to-one correspondence between the UAV flight trajectory and the building entity space, providing a unified spatial reference plane for contour extraction and deformation separation.

[0024] It should be noted that step S1, which performs control command parsing and multi-source data benchmark alignment, provides a unified spatial reference system for decoupling deformation interference. Conventional aerial image stitching processes lack a clear geographic benchmark mapping, relying directly on image feature matching for inter-frame stitching. In low-altitude surround scenarios during fire rescue, this can easily lead to global coordinate system inaccuracies due to accumulated initial pose drift. Step S1, by introducing a global spatial benchmark and performing coordinate system alignment processing, forcibly anchors the discretely acquired continuous aerial image frame sequence to the three-dimensional coordinate system of the building entity. The coordinate alignment sequence eliminates initial sensor errors and coordinate system axial deviations, ensuring that contour stability verification and spatial deviation quantification are performed within a unified reference system. The data preprocessing process prevents the transmission of initial pose errors to the deformation analysis stage, ensuring that thermal deformation boundary identification and pose compensation correction are based on an accurate geographic spatial benchmark, laying a data foundation for accurate cross-regional mapping of target annotation locations during thermal expansion.

[0025] Preferably, step S1 retrieves the initial pose parameters and global spatial reference of the continuous aerial frame sequence by parsing the aerial stitching control command and the target area identifier. The initial pose parameters and global spatial reference are then subjected to coordinate system alignment processing, which includes translation vector calculation and rotation matrix superposition, to generate a coordinate alignment sequence. The above processing path establishes a one-to-one correspondence between the UAV flight trajectory and the building entity space and eliminates the initial sensor error and coordinate system axial deviation. This ensures that contour stability verification and spatial deviation quantification are performed within a unified reference system, thereby blocking the transmission of initial pose error to the deformation analysis stage and ensuring that thermal deformation boundary identification and pose compensation correction are based on an accurate geospatial reference. This lays the data foundation for accurate cross-regional mapping of the target annotation position during the thermal expansion process.

[0026] S2: Perform contour stability verification on the continuous aerial frame sequence based on the coordinate alignment sequence, separate the deformation interference area from the fixed structure area, and generate deformation state parameters; In some embodiments, step S2 involves performing contour stability verification on the continuous aerial frame sequence based on the coordinate alignment sequence, separating the deformation interference region from the fixed structure region, and generating deformation state parameters, including steps S21 to S23: Step S21: Extract the edge contour lines of adjacent frames in the continuous aerial frame sequence, determine the curvature change trend and displacement gradient distribution between the edge contour lines of adjacent frames, and generate the contour evolution matrix.

[0027] It should be noted that in the fire rescue scenario of super high-rise buildings, the expansion and cracking of glass curtain walls due to heat causes the building edges to exhibit non-rigid deformation in continuous aerial frames. Conventional image stitching processes that directly extract the original edges and perform feature matching will misjudge the contour distortion caused by thermal deformation as drone pose drift. Step S21 transforms the messy edge pixels into a structured data set that records the evolution of local geometric shape by quantifying the curvature change trend and displacement gradient distribution, providing an orthogonal reference benchmark to distinguish between real pose changes and structural thermal deformation.

[0028] In some embodiments, step S21, which involves determining the curvature change trend and displacement gradient distribution between the edge contour lines of adjacent frames to generate a contour evolution matrix, includes steps S211 to S214: Step S211: Extract contour pixels along the edge contour lines of adjacent frames at a fixed step size, perform neighborhood smoothing on the contour pixels, and generate a smooth contour sequence.

[0029] Furthermore, the spatial distribution neighborhood of the contour pixels is obtained, and a neighborhood distribution ratio matrix matching the size of the spatial distribution neighborhood is constructed. The construction of the neighborhood distribution ratio matrix follows the distance decay principle, assigning a higher numerical ratio to neighborhood points closer to the contour pixels and a gradually decreasing numerical ratio to neighborhood points farther away from the contour pixels. The coordinate values ​​of the neighborhood distribution ratio matrix and the contour pixels are fused to generate smooth pixel coordinates. The smooth pixel coordinates are rearranged in order according to a fixed step size to obtain a smooth contour sequence. The fixed step size is set according to the matching relationship between the low-altitude flight altitude of the UAV and the field of view of the optical lens to ensure that the sampling density is sufficient to cover the fine scale of the curtain wall bulging. This processing filters out edge jitter caused by environmental airflow disturbances and retains the true physical contour of the building facade.

[0030] Step S212: Determine the contour direction of each pixel based on the smooth contour sequence, and convert the contour direction into a tangent direction vector.

[0031] Step S213: Generate the corresponding normal direction vector based on the tangent direction vector.

[0032] It is easy to understand that the tangent direction vector reflects the extension trend of the contour in the current frame, while the normal direction vector reflects the vertical offset direction of the contour bulging outward or collapsing inward. The operation of generating the normal direction vector performs a vertical rotation transformation on the tangent direction vector in the two-dimensional image plane to obtain the normal direction vector perpendicular to the contour tangent.

[0033] Step S214: Combine the tangent direction vector and the normal direction vector to generate the contour evolution matrix.

[0034] The contour evolution matrix fully records the local curvature offset and normal displacement information of the building boundary between adjacent frames, avoiding the problem of ignoring the deformation direction in a single coordinate comparison.

[0035] Step S22: Compare the inter-frame displacement difference of each contour region in the contour evolution matrix, accumulate the inter-frame displacement difference along the frame sequence direction to obtain the cumulative displacement value, set the upper limit threshold and the lower limit threshold of the reference vibration, and divide the contour region into deformation interference region, fixed structure region or invalid occlusion region based on the comparison relationship between the cumulative displacement value and the threshold interval.

[0036] It should be noted that the building facade exhibits a complex spatial response state under fire conditions, and a single judgment logic cannot cover all situations. The upper limit threshold of the baseline vibration is set according to the critical point of the initial deformation of the glass curtain wall under heat, and the lower limit threshold of the baseline vibration is set according to the allowable sway amplitude of the super high-rise building under conventional wind load. The upper limit threshold of the baseline vibration is greater than the lower limit threshold of the baseline vibration. The classification rules cover all numerical ranges and endpoint states to ensure that no judgment is missed.

[0037] In some embodiments, step S22, which divides the contour region into a deformation interference region, a fixed structure region, or an invalid occlusion region based on the comparison between the cumulative displacement value and the threshold interval, includes steps S221 to S223: Step S221: When the cumulative displacement value is greater than or equal to the upper limit threshold of the reference vibration, the contour area is divided into the deformation interference area.

[0038] It should be noted that the bulging and crack propagation of the curtain wall under high temperature is manifested as a continuous increase in displacement in one direction along the frame sequence. The cumulative displacement value exceeds the upper limit threshold of the reference vibration, indicating that the contour area has undergone substantial physical migration. Classifying the contour area as a deformation interference area makes the thermally induced deformation boundary enter an independent deviation quantization channel and no longer participate in the global pose calculation, directly cutting off the path of deformation interference to the splicing coordinate system.

[0039] Step S222: When the cumulative displacement value is less than the upper limit threshold of the reference vibration and greater than or equal to the lower limit threshold of the reference vibration, the processing system performs a frame sequence period check operation. The contour area where the cumulative displacement value shows a reciprocating zero state or maintains a stable fluctuation between the upper limit threshold of the reference vibration and the lower limit threshold of the reference vibration is divided into a fixed structure area.

[0040] It is easy to understand that load-bearing walls or the outer wall of the core tube that are not directly affected by high temperatures are only affected by environmental wind loads and their displacement alternates periodically within the threshold range. Classifying the contour area as a fixed structural area provides a real spatial reference for pose calculation. Traditional splicing processes treat wind vibration and micro-motion as static features, which can easily accumulate calculation errors. This solution confirms the structural stability through periodic verification to ensure that the fixed structural area maintains geometric anchoring during splicing mapping.

[0041] Step S223: When the cumulative displacement value is less than the lower limit threshold of the reference vibration or when the cumulative displacement value exhibits disordered jumping state and exceeds the effective range of sensor acquisition, the contour area is marked as an invalid occlusion area and is removed.

[0042] It should be noted that dense smoke or lens glare can cause edge features to break and displacement data to lose physical continuity. If the cumulative displacement value is lower than the baseline vibration lower limit threshold or exhibits disordered jumps, it means that the image data can no longer reflect the true building outline. The outline area is marked as an invalid occlusion area and directly removed to avoid the broken data being misidentified as a pose change. The classification logic fully covers all situations, including those greater than the upper limit range, those within the threshold range, those less than the lower limit range, and data anomalies, and the endpoint determination is clear to prevent classification blind spots from causing the stitching boundary to shift.

[0043] Step S23: Traverse the boundary pixels covered by the deformation interference area, obtain the offset distance and shape distortion of the boundary pixels between adjacent frames, and combine them to generate deformation state parameters.

[0044] Specifically, the offset distance is obtained by measuring the spatial interval between the corresponding positions of the boundary pixels in adjacent frames, and the shape distortion is obtained by comparing the difference in the local contour curvature values ​​of the boundary pixels in adjacent frames. The offset distance and shape distortion are aligned according to the pixel index order and combined to generate deformation state parameters. The deformation state parameters completely record the migration trajectory and shape distortion degree of the building boundary under high temperature.

[0045] It should be noted that step S2 performs contour stability verification and region classification to directly solve the core technical problem of conventional aerial stitching misidentifying thermally deformed boundaries as drone pose drift. Traditional stitching processes assume that building facades are rigid fixed structures, and in fire rescue scenarios, the bulging of glass curtain walls and smoke obstruction cause local migration of continuous frame boundaries. If the migrated boundaries are directly included in the global homography mapping, the stitching mesh will be pulled by thermal expansion, resulting in linkage offset, and the target annotation coordinates will be misaligned across regions.

[0046] Preferably, step S2 quantifies the local geometric evolution by constructing a contour evolution matrix and physically isolates the unidirectional cumulative thermal deformation, periodic wind vibration and swaying and disordered smoke flow obstruction based on the complete comparison of the cumulative displacement value and the dual threshold interval. The fixed structure area maintains the stability of the global coordinate system and the deformation interference area enters the independent deviation channel. This processing method avoids deformation interference from polluting the global stitching mesh and makes the UAV's real pose calculation no longer constrained by the thermal expansion of the building. This provides a reliable data basis for the accurate anchoring of the target annotation position during the thermal expansion process and provides accurate interference source feature input for the spatial deviation calculation in step S3.

[0047] S3: Determine the spatial deviation based on the deformation state parameters, generate the region association weight based on the spatial deviation, and determine the pose compensation parameters based on the region association weight. In some embodiments, step S3, which involves determining the spatial deviation based on the deformation state parameters, generating region association weights based on the spatial deviation, and determining pose compensation parameters based on the region association weights, includes steps S31 to S33: Step S31: Map the deformation state parameters to the corresponding pixel grid of the continuous aerial frame sequence, obtain the spatial displacement data between the corresponding pixel grid and the fixed structure area, and generate the local spatial deviation.

[0048] It should be noted that the deformation state parameters record the offset distance and morphological distortion of the boundary pixels. Step S31 performs spatial interpolation on the discrete boundary parameters to the two-dimensional image plane, so that each grid node in the corresponding pixel grid of the continuous aerial frame sequence has a clear displacement vector. The fixed structure region is used as the geometric reference origin. The spatial displacement data is obtained by comparing the current position of the corresponding pixel grid with the reference position of the fixed structure region. The local spatial deviation thus fully characterizes the actual distance and direction of each grid node in the continuous aerial frame sequence from the original static reference. This mapping operation extends the one-dimensional boundary features to the entire image domain to provide continuous data support for weight allocation.

[0049] Step S32: Determine the attenuation coefficient field based on the modulus distribution of the local spatial deviation, and perform spatial overlay processing on the mapping confidence of the attenuation coefficient field and the corresponding pixel grid to generate the region association weight.

[0050] In some embodiments, step S32, which involves determining the attenuation coefficient field based on the modulus distribution of the local spatial deviation, and performing spatial overlay processing on the mapping confidence of the attenuation coefficient field and the corresponding pixel grid to generate region association weights, includes steps S321 to S323: Step S321: Statistically analyze the distribution of extreme points of local spatial deviation in the global pixel grid to determine the deformation diffusion center and deformation static boundary.

[0051] It should be noted that the modulus distribution of local spatial deviation reflects the radiation range of thermally induced deformation on the building facade, and the extreme point distribution refers to the set of pixel nodes where the modulus of local spatial deviation reaches its peak value. The processing system marks the geometric center of the region where the modulus peak value is concentrated as the deformation diffusion center, and marks the line connecting the pixel nodes where the modulus drops to the environmental noise baseline as the deformation static boundary. The deformation diffusion center corresponds to the area where the glass curtain wall expands most violently due to heat, and the deformation static boundary corresponds to the outer edge of the stable structure that has not yet been affected by thermal stress. Clearly defining the spatial positions of the two provides a geometric reference for the spatial distribution of the attenuation coefficient.

[0052] Step S322: Determine the radial attenuation field starting from the deformation diffusion center, and generate the attenuation coefficient based on the distribution relationship between the relative distance from the global pixel grid to the deformation diffusion center and the length of the deformation static boundary.

[0053] Specifically, the relative distance from each grid node in the global pixel grid to the deformation diffusion center is calculated, the total length of the deformation static boundary is obtained, the relative distance and the length of the deformation static boundary are proportionally allocated to obtain the distance ratio value, and the inverse mapping is performed based on the distance ratio value to generate the attenuation coefficient corresponding to each grid node.

[0054] It should be noted that the influence range of thermally induced deformation gradually weakens with increasing distance from the deformation diffusion center. The attenuation coefficient must cover all spatial locations and exhibit a continuous transition. When the relative distance is equal to zero, the attenuation coefficient is set to the lowest baseline value to ensure that the mapping confidence of the deformation core region is reduced to the minimum. When the relative distance is greater than zero and less than the deformation static boundary length, the attenuation coefficient increases linearly with the relative distance to reflect the physical law of thermal stress gradually attenuating towards the periphery. When the relative distance is equal to the deformation static boundary length, the attenuation coefficient reaches the highest baseline value to mark the complete disappearance of the deformation influence. When the relative distance is greater than the deformation static boundary length, the attenuation coefficient remains unchanged at the highest baseline value to ensure that the unaffected areas maintain the original splicing accuracy. This classification rule covers the entire distance range and endpoint state to avoid weight jumps that cause tearing and misalignment of the splicing mesh.

[0055] Step S323: Perform spatial domain fusion processing on the mapping confidence of the attenuation coefficient and the corresponding pixel grid, and output the region association weight.

[0056] In some embodiments, step S323 involves performing spatial fusion processing on the mapping confidence of the attenuation coefficient and the corresponding pixel grid to output the region association weight, including steps S3231 to S3233: Step S3231: Construct a spatial weight distribution array that matches the size of the continuous aerial frame sequence, and fill the central region of the spatial weight distribution array with the attenuation coefficients according to the spatial distribution.

[0057] It should be noted that continuous aerial frame sequence size matching means that the number of rows and columns of the spatial weight distribution array is completely consistent with the pixel resolution of the continuous aerial frame sequence. The spatial weight distribution array, as a data carrier, has its central area corresponding to the location of the deformation diffusion center. The attenuation coefficients are filled one-to-one according to the spatial coordinates so that the numerical distribution of the spatial weight distribution array completely coincides with the thermal deformation radiation pattern of the building facade. This filling operation establishes a physical correspondence between the attenuation coefficients and image pixels to avoid spatial misalignment that leads to weight distribution distortion.

[0058] Step S3232: Perform feature overlay processing on the mapping confidence and the spatial weight distribution matrix to generate a local weight distribution matrix.

[0059] It is easy to understand that the mapping confidence reflects the degree of trust that the original stitching algorithm has in the local features of the image. The feature overlay process performs a bit-by-bit proportional fusion of the attenuation coefficient in the spatial weight distribution matrix with the mapping confidence. The attenuation coefficient value corresponding to the thermally deformed region is lower, so the mapping confidence of the corresponding grid node is actively suppressed after the feature overlay process. The attenuation coefficient value corresponding to the fixed structure region is higher, so the mapping confidence remains at the original level after the feature overlay process. The local weight distribution matrix thus achieves the synchronous execution of dynamic interference reduction and static benchmark preservation.

[0060] Step S3233: Perform neighborhood numerical smoothing on the local weight distribution matrix according to the sliding window range, and output the region correlation weights.

[0061] Specifically, the coverage size and movement step size of the sliding window are set. The coverage size of the sliding window is set according to the matching relationship between the image resolution and lens distortion radius of the low-altitude aerial photography of the UAV. The movement step size is set according to an integer fraction of the coverage size of the sliding window to ensure that the window movement trajectory covers all grid nodes. The sliding window performs row-by-row scanning along the row and column directions of the local weight distribution matrix. When the sliding window is in the edge area of ​​the image and part of the window range exceeds the boundary of the local weight distribution matrix, the system will mirror and copy the area beyond the boundary to the internal effective area to keep the number of nodes covered by the window constant. The values ​​of each grid node in the window coverage area are calculated by weighted averaging to obtain the smooth value of the center node. After traversing all grid nodes, the region association weight is output.

[0062] It should be noted that during the feature overlay process, the local weight distribution matrix may cause abrupt changes in the values ​​of adjacent nodes due to pixel-level discrete calculations. The sliding window range performs neighborhood numerical smoothing to eliminate high-frequency numerical jumps and make the regional associated weights present a continuous and gradual change in space. The edge mirroring operation ensures that the boundary grid nodes obtain complete neighborhood references to prevent weight truncation at the splicing boundary. This processing ensures that the regional associated weights transition smoothly in the global coordinate system to avoid mesh distortion caused by abrupt changes in local weights.

[0063] Step S33: Perform a unified numerical range processing on the region association weights, and determine the pose compensation parameters by combining the rigid transformation parameters of the fixed structure region.

[0064] Specifically, the minimum and maximum values ​​in the regional association weights are obtained, and a unified target interval lower limit and a unified target interval upper limit are set. The unified target interval lower limit and the unified target interval upper limit are set according to the pose calculation data format requirements of the UAV flight control system. The values ​​of each node in the regional association weights are mapped linearly to the unified target interval lower limit and the unified target interval upper limit to complete the value range unification processing. The unified regional association weights and the rigid transformation parameters of the fixed structure region are subjected to matrix fusion operation to output pose compensation parameters. The pose compensation parameters are used to correct the actual flight trajectory of the UAV in the global space.

[0065] It should be noted that rigid transformation parameters refer to pose transformation data of a fixed structural region that maintains its geometric shape during spatial mapping, encompassing spatial translation vectors and spatial rotation matrices. The processing system extracts rigid transformation parameters based on the contour features of the fixed structural region. These parameters reflect the true spatial pose of the building's load-bearing components, unaffected by thermal deformation, in continuous aerial frames. When determining pose compensation parameters, the uniformly processed region association weights are fused with the rigid transformation parameters using a matrix operation, directly coupling the weighting effect of thermal deformation interference regions with the stable pose of the fixed structural region. The rigid transformation parameters provide a geometric reference unaffected by thermal expansion, ensuring that pose compensation calculations are always anchored to the true static building contour, preventing the compensation direction from deviating from the actual flight trajectory.

[0066] It should be noted that after the region association weights are smoothed, there may be a mismatch between the numerical distribution range and the input requirements of the pose calculation module. The numerical range is uniformly processed to compress the weight data into a standard range that the flight control system can recognize, so as to ensure that the compensation parameters are directly connected to the navigation calculation link. Combined with the rigid transformation parameters of the fixed structure region, matrix fusion is performed so that the weight reduction effect of the deformation interference region is directly converted into the translation and rotation compensation of the pose coordinates. This step completes the final mapping from image domain weight allocation to pose domain coordinate correction.

[0067] Preferably, step S3 performs spatial deviation quantization and regional correlation weight allocation, directly addressing the core technical problem in conventional aerial image stitching where thermally induced deformation boundaries continuously participate in spatial mapping, leading to global stitching boundary linkage shifts. Addressing the technical flaw in traditional image stitching processes where all image features are included in equal-weighted homography transformation, causing local migrations resulting from thermal expansion and cracking of glass curtain walls to be directly accumulated in the pose calculation matrix and leading to cross-regional misalignment of target annotation positions along false boundaries, this solution determines the deformation diffusion center and deformation static boundary through statistical extreme point distribution to construct a radial attenuation field covering the entire distance interval and endpoint states, thus enabling... The attenuation coefficient and mapping confidence are spatially fused to actively suppress the mapping confidence of the thermally deformed region and maintain high positional confidence in the fixed structure region. The local weight distribution matrix is ​​smoothed by the neighborhood of the sliding window and outputs the region-related weights to eliminate the risk of tearing of the spliced ​​mesh by numerical jumps. Finally, the pose compensation parameters are generated by fusion with the rigid transformation parameters after the numerical range is unified. The above technical processing cuts off the pulling effect of thermal deformation on the global coordinate system and makes the UAV's true pose calculation return to the fixed structure reference, thereby ensuring that the target annotation position no longer undergoes cross-regional misalignment during thermal expansion and providing accurate compensation input for the global mapping correction in step S4.

[0068] In some embodiments, step S4 involves performing a global mapping correction on the initial pose parameters based on the pose compensation parameters, generating a corrected stitching sequence, and outputting annotation location information based on the corrected stitching sequence and the target region identifier, including steps S41 to S43: Step S41: Decompose the pose compensation parameters into translation compensation components and rotation compensation components, and perform spatial transformation processing with the initial pose parameters respectively to generate the corrected pose sequence.

[0069] It should be noted that during the low-altitude circling flight of the drone at the fire scene, the disturbance of thermal airflow and building deformation caused the initial pose parameters to contain spurious drift components. Step S41 decomposes the pose compensation parameters into independent spatial translation and spatial rotation components to eliminate the cross-effect of coupling interference on a single coordinate axis. The translation compensation component corrects the drone's positional offset in three-dimensional space, and the rotation compensation component corrects the pitch and yaw deviations of the shooting perspective. The translation and rotation compensation components are then subjected to spatial transformation processing with the initial pose parameters to generate a corrected pose sequence. This processing path removes the coordinate system distortion introduced by thermal deformation, restoring the geometric consistency of the drone's true flight trajectory in the spatial dimension.

[0070] Step S42: Perform grid deformation coordinate mapping processing on the continuous aerial frame sequence according to the corrected pose sequence to generate a corrected stitched sequence.

[0071] In some embodiments, step S42, which involves performing mesh deformation coordinate mapping processing on the continuous aerial frame sequence based on the corrected pose sequence to generate a corrected stitched sequence, includes steps S421 to S424: Step S421: Perform mesh generation on the continuous aerial frame sequence according to the corrected pose sequence to generate an initial stitched mesh.

[0072] It should be noted that global homography transformation cannot handle localized thermal deformation of building facades. Mesh generation divides the continuous aerial frame sequence into independent mapping units. Based on the spatial topology provided by the corrected pose sequence, a regular grid node array is constructed on the pixel plane of the continuous aerial frame sequence to generate an initial stitched mesh. The mesh density of the initial stitched mesh is set according to the image resolution and the distribution range of the deformation region, ensuring that each mesh unit covers a single physical structural surface or a single deformation transition zone. This mesh generation operation provides an independently adjustable geometric framework for local deformation compensation.

[0073] Step S422: Map the corrected pose sequence to the initial stitched mesh and determine the global projection error of the initial stitched mesh.

[0074] Specifically, the camera intrinsic and extrinsic parameters from the corrected pose sequence are injected node by node into the initial stitched mesh. The positional deviation of each mesh vertex between the theoretical imaging plane and the actual pixel plane is calculated to generate a global projection error. The global projection error value reflects the degree of matching between the corrected pose sequence and the pixel distribution of the current frame.

[0075] It should be noted that thermal deformation causes bulges or depressions on the building surface in three-dimensional space, inevitably resulting in a deviation between the theoretical projected coordinates and the actual pixel positions. By injecting parameters node by node and calculating the positional deviation, the processing system accurately quantifies the residual deformation of each mesh unit, providing a clear correction target for vertex displacement adjustment.

[0076] Step S423: Perform mesh vertex displacement adjustment processing according to the global projection error distribution to generate a deformation compensation mesh.

[0077] Specifically, an error convergence benchmark is set. If the global projection error is greater than the benchmark, the processing system performs displacement iterations on the corresponding mesh vertex along the opposite direction of the deviation vector. If the global projection error is less than or equal to the benchmark, the processing system keeps the coordinates of the corresponding mesh vertex unchanged. After traversing all mesh vertices in the initial stitched mesh, the deformation compensation mesh is output.

[0078] It should be noted that the mesh vertex displacement adjustment process must cover the entire error range to ensure stitching continuity. Regions with errors greater than the convergence benchmark correspond to areas of severe thermal deformation or transitional zones blocked by smoke flow; moving vertices in the opposite direction of the deviation can offset projection distortion. Regions with errors less than or equal to the convergence benchmark correspond to fixed structure regions or stable regions that have already been compensated; keeping vertex coordinates unchanged avoids introducing mesh tearing due to overcorrection. Endpoint states are clearly defined, and displacement iteration automatically terminates when convergence conditions are met, ensuring a smooth transition of the deformation-compensated mesh globally without abrupt breaks.

[0079] Step S424: Resample the pixel content of the continuous aerial frame sequence into the corresponding topology of the deformation compensation grid, and output the corrected stitching sequence.

[0080] It's easy to understand that the deformation compensation mesh only contains optimized spatial vertex coordinates, requiring the original image pixels to be remapped to the new mesh. The system reads the original pixel array from a continuous sequence of aerial images, constructs a mapping topology based on the vertex connections of the deformation compensation mesh, performs pixel color value filling operations to fill the pixel data into the new mesh cells, and outputs a corrected stitching sequence. This resampling operation eliminates image stretching and compression artifacts caused by local thermal expansion, restoring the physical proportions of the building facade texture in the stitched result.

[0081] Step S43: Transform the pixel coordinates of the corrected stitched sequence to a global unified coordinate system, determine the corresponding annotation position information based on the target area identifier, and output the annotation position information.

[0082] Specifically, the boundary of the region of interest corresponding to the target region identifier is read. Then, the pixel coordinates within the boundary of the region of interest in the corrected stitching sequence are subjected to an affine transformation to convert them to geospatial coordinates in a globally unified coordinate system. Based on the converted geospatial coordinates, annotation location information is generated and output.

[0083] It should be noted that traditional stitching results directly output pixel coordinates, which lack actual geographical reference at the fire rescue site. Step S43 corrects the pixel coordinates of the stitching sequence to a globally unified coordinate system, so that the labeled location information directly corresponds to the actual floor location or facade zoning of the high-rise building. The target area identifier limits the labeled range, avoiding the waste of computational resources on coordinate transformation of non-key areas. The output results can be directly connected to the fire command terminal to achieve accurate positioning of the rescue target in real three-dimensional space.

[0084] It should be noted that step S4 performs global mapping correction and annotation coordinate output, which directly solves the core technical problem of cross-regional misalignment of target annotation positions during thermal expansion in conventional aerial photography stitching. Traditional stitching processes only perform simple pixel overlay after generating panoramic images. The continuous thermal deformation of the building facade causes the stitching mesh to accumulate geometric distortion, and the target annotation coordinates drift spatially with the false boundary.

[0085] Preferably, step S4 eliminates coordinate system cross-interference by decomposing the pose compensation parameters into translational and rotational compensation components, generates an initial stitched mesh based on the corrected pose sequence, quantizes the global projection error node by node, performs mesh vertex displacement adjustment processing based on a complete comparison of the global projection error distribution and the error convergence benchmark value to generate a deformation compensation mesh, resamples the pixel content of the continuous aerial frame sequence into the corresponding topology of the deformation compensation mesh, outputs the corrected stitched sequence, and finally transforms the pixel coordinates of the corrected stitched sequence to a global unified coordinate system and outputs the annotation position information in conjunction with the target area identifier. The above technical processing cuts off the traction effect of thermal deformation on the global stitched mesh and makes the actual pose calculation of the UAV return to the fixed structural benchmark, thereby ensuring that the target annotation position no longer undergoes cross-regional misalignment during thermal expansion and providing accurate spatial positioning basis for fire rescue command.

[0086] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0087] Based on the same inventive concept, this application also provides a real-time stitching and target annotation system for UAV aerial images. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the real-time stitching and target annotation system for UAV aerial images provided below can be found in the limitations of the real-time stitching and target annotation method for UAV aerial images described above, and will not be repeated here.

[0088] In one exemplary embodiment, such as Figure 3 As shown, a real-time stitching and target annotation system for UAV aerial images is provided, including: The alignment module receives aerial photography stitching control commands and target area identifiers, calls the initial pose parameters and global spatial reference of the continuous aerial photography frame sequence, performs coordinate system alignment processing on the initial pose parameters and global spatial reference, and generates a coordinate alignment sequence. The verification module performs contour stability verification on a continuous aerial frame sequence based on the coordinate alignment sequence, separates the deformation interference area from the fixed structure area, and generates deformation state parameters. The compensation module determines the spatial deviation based on the deformation state parameters, generates regional association weights based on the spatial deviation, and determines the pose compensation parameters based on the regional association weights. The correction module performs global mapping correction on the initial pose parameters based on the pose compensation parameters, generates a corrected stitching sequence, and outputs annotation position information based on the corrected stitching sequence and the target region identifier.

[0089] The modules in the aforementioned real-time stitching and target annotation system for UAV aerial images can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0090] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational 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 input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements a method for real-time stitching and target annotation of UAV aerial images. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0091] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0092] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0093] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0094] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0095] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0096] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0097] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0098] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for real-time stitching and target annotation of drone aerial images, characterized in that, include: Receive aerial photography stitching control instructions and target area identifiers, call the initial pose parameters and global spatial reference of the continuous aerial photography frame sequence, perform coordinate system alignment processing on the initial pose parameters and the global spatial reference, and generate a coordinate alignment sequence; Based on the coordinate alignment sequence, the contour stability of the continuous aerial frame sequence is checked to separate the deformation interference region from the fixed structure region and generate deformation state parameters. Based on the deformation state parameters, the spatial deviation is determined, the regional association weight is generated based on the spatial deviation, and the pose compensation parameters are determined based on the regional association weight. The initial pose parameters are globally mapped and corrected according to the pose compensation parameters to generate a corrected stitching sequence. The annotation position information is then output based on the corrected stitching sequence and the target region identifier.

2. The method for real-time stitching and target annotation of UAV aerial images according to claim 1, characterized in that, The step of performing contour stability verification on the continuous aerial frame sequence based on the coordinate alignment sequence, separating the deformation interference region from the fixed structure region, and generating deformation state parameters includes: Extract the edge contour lines of adjacent frames in the continuous aerial frame sequence, determine the curvature change trend and displacement gradient distribution between the edge contour lines of adjacent frames, and generate a contour evolution matrix; By comparing the inter-frame displacement differences of each contour region in the contour evolution matrix, the contour regions in which the displacement differences are unidirectionally accumulated along the frame order direction are classified as deformation interference regions, and the contour regions in which the displacement differences are constant along the frame order direction are classified as fixed structure regions. The boundary pixels contained in the deformation interference region are traversed, and the offset distance and morphological distortion of the boundary pixels between adjacent frames are obtained. The deformation state parameters are then generated by combining these parameters.

3. The method for real-time stitching and target annotation of UAV aerial images according to claim 1, characterized in that, The process of determining the spatial deviation based on the deformation state parameters, generating region association weights based on the spatial deviation, and determining pose compensation parameters based on the region association weights includes: The deformation state parameters are mapped to the corresponding pixel grid of the continuous aerial frame sequence, and the spatial displacement data between the corresponding pixel grid and the fixed structure region is obtained to generate local spatial deviation. The attenuation coefficient field is determined based on the modulus distribution of the local spatial deviation. The attenuation coefficient field and the mapping confidence of the corresponding pixel grid are subjected to spatial superposition processing to generate region association weights. The numerical range of the associated weights of the region is uniformly processed, and the pose compensation parameters are determined by combining the rigid transformation parameters of the fixed structural region.

4. The method for real-time stitching and target annotation of UAV aerial images according to claim 2, characterized in that, The process of determining the curvature change trend and displacement gradient distribution between the edge contour lines of adjacent frames and generating a contour evolution matrix includes: Extract contour pixels along the edge contour lines of the adjacent frames at a fixed step size, perform neighborhood smoothing on the contour pixels, and generate a smooth contour sequence. The contour direction of each pixel is determined according to the smooth contour sequence, the contour direction is converted into a tangent direction vector, and a corresponding normal direction vector is generated based on the tangent direction vector. The contour evolution matrix is ​​generated by combining the tangent direction vector and the normal direction vector.

5. The method for real-time stitching and target annotation of UAV aerial images according to claim 3, characterized in that, The step of determining the attenuation coefficient field based on the modulus distribution of the local spatial deviation, and performing spatial superposition processing on the mapping confidence of the attenuation coefficient field and the corresponding pixel grid to generate region association weights includes: The distribution of extreme points of the local spatial deviation in the global pixel grid is statistically analyzed to determine the deformation diffusion center and the deformation static boundary. A radial attenuation field is determined starting from the deformation diffusion center, and an attenuation coefficient is generated based on the distribution relationship between the relative distance from the global pixel grid to the deformation diffusion center and the length of the deformation static boundary. The attenuation coefficient and the mapping confidence of the corresponding pixel grid are subjected to spatial domain fusion processing to output the region association weight.

6. The method for real-time stitching and target annotation of UAV aerial images according to claim 1, characterized in that, The step of performing global mapping correction on the initial pose parameters according to the pose compensation parameters, generating a corrected stitching sequence, and outputting annotation location information according to the corrected stitching sequence and the target region identifier includes: The pose compensation parameters are decomposed into translation compensation components and rotation compensation components, and spatial transformation processing is performed on the initial pose parameters respectively to generate a corrected pose sequence. Based on the corrected pose sequence, perform grid deformation coordinate mapping processing on the continuous aerial frame sequence to generate a corrected stitched sequence; The pixel coordinates of the corrected stitched sequence are transformed to a globally unified coordinate system, and the corresponding annotation position information is determined according to the target region identifier, and the annotation position information is output.

7. The method for real-time stitching and target annotation of UAV aerial images according to claim 4, characterized in that, The step of performing neighborhood smoothing on the contour pixels to generate a smooth contour sequence includes: Obtain the spatial distribution neighborhood of the contour pixels, and construct a neighborhood distribution ratio matrix that matches the size of the spatial distribution neighborhood; The neighborhood distribution ratio matrix and the coordinate values ​​of the contour pixels are weighted and fused to generate smooth pixel coordinates. The smooth pixel coordinates are rearranged in order according to the fixed step size to obtain a smooth contour sequence.

8. The method for real-time stitching and target annotation of UAV aerial images according to claim 5, characterized in that, The step of performing spatial domain fusion processing on the mapping confidence of the attenuation coefficient and the corresponding pixel grid, and outputting the region association weight, includes: Construct a spatial weight distribution array that matches the size of the continuous aerial frame sequence, and fill the central region of the spatial weight distribution array with the attenuation coefficients according to the spatial distribution; The mapping confidence is combined with the spatial weight distribution matrix to perform feature superposition processing, thereby generating a local weight distribution matrix; The local weight distribution matrix is ​​subjected to neighborhood numerical smoothing based on the sliding window range, and the region-related weights are output.

9. The method for real-time stitching and target annotation of UAV aerial images according to claim 6, characterized in that, The step of performing mesh deformation coordinate mapping processing on the continuous aerial frame sequence based on the corrected pose sequence to generate a corrected stitched sequence includes: The continuous aerial frame sequence is subjected to mesh division processing based on the corrected pose sequence to generate an initial stitched mesh. The corrected pose sequence is mapped to the initial stitched mesh, and the global projection error of the initial stitched mesh is determined. Based on the global projection error distribution, perform mesh vertex displacement adjustment processing to generate a deformation compensation mesh; The pixel content of the continuous aerial frame sequence is resampled into the corresponding topology of the deformation compensation grid, and the corrected stitching sequence is output.

10. A real-time stitching and target annotation system for UAV aerial images, employing the real-time stitching and target annotation method for UAV aerial images as described in any one of claims 1 to 9, characterized in that, include: The alignment module receives aerial photography stitching control instructions and target area identifiers, calls the initial pose parameters and global spatial reference of the continuous aerial photography frame sequence, performs coordinate system alignment processing on the initial pose parameters and the global spatial reference, and generates a coordinate alignment sequence. The verification module performs contour stability verification on the continuous aerial frame sequence based on the coordinate alignment sequence, separates the deformation interference region from the fixed structure region, and generates deformation state parameters. The compensation module determines the spatial deviation based on the deformation state parameters, generates regional association weights based on the spatial deviation, and determines pose compensation parameters based on the regional association weights. The correction module performs global mapping correction on the initial pose parameters according to the pose compensation parameters, generates a corrected stitching sequence, and outputs annotation position information based on the corrected stitching sequence and the target region identifier.

Images