Search

CN-122023450-A - Unmanned plane edge calculation landslide segmentation modeling method based on dynamic attitude correction

CN122023450ACN 122023450 ACN122023450 ACN 122023450ACN-122023450-A

Abstract

The invention relates to the technical field of image processing, and discloses an unmanned plane edge calculation landslide segmentation modeling method based on dynamic gesture correction, which comprises the steps of acquiring an original image acquired by an unmanned plane, synchronously acquiring multidimensional gesture data to acquire a complete gesture data set, extracting dynamic characteristic calculation gesture fluctuation indexes to establish a linkage mapping rule, and resolving deformation parameters to obtain coordinate correction coefficients so as to correct the original image, preprocessing and determining a primary contour range, determining a disaster core area according to an attitude fluctuation index judging state, triggering closed loop calibration and reconstruction of the coordinate correction coefficients when the vision and physical deformation translation vector difference value presents an increasing trend, and evaluating boundary stability to generate landslide segmentation results. The invention overcomes deformation distortion through closed loop calibration of multidimensional gesture and visual translation, improves the accuracy of boundary space, establishes linkage mapping rule switching workflow to realize calculation force distribution as required, and fits unstable boundary differentiation to eliminate saw-tooth distortion, so that the segmentation result has high boundary closure degree.

Inventors

  • JIA YANG
  • XIANG BO
  • DING YULIN
  • WANG ZHONGWEN

Assignees

  • 四川省公路规划勘察设计研究院有限公司
  • 西南交通大学

Dates

Publication Date
20260512
Application Date
20260410

Claims (10)

  1. 1. The unmanned aerial vehicle edge calculation landslide segmentation modeling method based on dynamic attitude correction is characterized by comprising the following steps of: acquiring an original image acquired by an unmanned aerial vehicle, synchronously acquiring multidimensional gesture data, and carrying out fusion processing on the multidimensional gesture data based on a dynamic fusion model to acquire a complete gesture data set containing gesture description parameters; extracting dynamic characteristics of the complete gesture data set to calculate gesture fluctuation indexes, and establishing a linkage mapping rule of the gesture fluctuation indexes and a computational power scheduling branch; According to the gesture description parameters, resolving deformation parameters comprising physical deformation translation vectors, acquiring coordinate correction coefficients based on the deformation parameters and combining the gesture fluctuation indexes, and correcting the original image by using the coordinate correction coefficients to generate corrected image data; preprocessing the corrected image data, extracting and screening edge lines to determine a preliminary contour range; Judging that the power calculation scheduling branch is in a stable state or a fluctuation state according to the gesture fluctuation index according to the linkage mapping rule, executing position updating in the stable state, and triggering a classification mechanism in the primary contour range in the fluctuation state to determine a disaster core area; extracting visual translation amounts of multiple frames of the corrected image data in the disaster core area, triggering closed-loop calibration to reconstruct the coordinate correction coefficients when the vector difference value of the visual translation amounts and the physical deformation translation vectors shows an increasing trend, and outputting dynamic boundary descriptions; And extracting the geometric element data described by the dynamic boundary, evaluating the stability of the local boundary, dividing the geometric element data into a stable data set and an unstable data subset according to the stability, and recombining the unstable data subset and the stable data set to generate a landslide segmentation result.
  2. 2. The unmanned aerial vehicle edge calculation landslide segmentation modeling method based on dynamic gesture correction of claim 1, wherein the fusion processing of the multidimensional gesture data based on the dynamic fusion model to obtain a complete gesture data set containing gesture description parameters specifically comprises: acquiring data covering height information, pitch angle, roll angle and yaw direction as the multidimensional gesture data; In a preset sliding time window, sequencing the multidimensional gesture data by adopting a median filter, and selecting a median value as denoising output to obtain a denoising gesture data set; performing linear normalization mapping on the extreme value range of the denoising gesture data set in a preset observation window to determine normalized gesture data; The dynamic fusion model comprising a system state transition matrix, a measurement noise covariance matrix and an error covariance matrix is constructed, a Kalman filtering mechanism is applied to execute state prediction and measurement updating on the normalized attitude data, kalman gain fusion real-time sampling values are calculated to output fusion attitude characteristic data, the fusion attitude characteristic data are stored in a circulation buffer zone to generate the complete attitude data set, and the pitching angle and the rolling angle are extracted to serve as the attitude description parameters.
  3. 3. The unmanned aerial vehicle edge calculation landslide segmentation modeling method based on dynamic gesture correction according to claim 2, wherein extracting the dynamic characteristics of the complete gesture data set to calculate a gesture fluctuation index, and establishing a linkage mapping rule of the gesture fluctuation index and a computational power scheduling branch specifically comprises: Extracting a three-dimensional acceleration vector and a three-dimensional angular velocity vector from the complete gesture data set as the dynamic characteristics; Calculating the innovation deviation between the observation vector of the dynamic characteristics and the prediction vector of the state prediction link in the dynamic fusion model, and smoothing the dynamic characteristics by adopting a weighted average method when the absolute value of the innovation deviation exceeds a preset dynamic deviation threshold; Extracting a numerical gradient of the dynamics characteristic after the smoothing treatment between continuous time frames, and calculating the attitude fluctuation index based on a preset proportional adjustment coefficient combined with a modulus value and a statistical variance of the numerical gradient; And when the gesture fluctuation index is larger than or equal to the scheduling stability threshold, mapping the calculated force scheduling branch to a full-quantization analysis branch, judging the calculated force scheduling branch to be in the fluctuation state, and establishing the linkage mapping rule.
  4. 4. The unmanned aerial vehicle edge calculation landslide segmentation modeling method based on dynamic gesture correction according to claim 2, wherein calculating deformation parameters including physical deformation translation vectors according to the gesture description parameters, acquiring coordinate correction coefficients based on the deformation parameters in combination with the gesture fluctuation index, and correcting the original image by using the coordinate correction coefficients to generate corrected image data specifically comprises: calculating a three-dimensional space rotation matrix by combining a camera internal reference matrix calibrated in advance and angle data contained in the gesture description parameters, and acquiring an image deformation matrix according to the three-dimensional space rotation matrix; Extracting a deformation scaling factor and the physical deformation translation vector from the image deformation matrix, and normalizing the deformation scaling factor and the physical deformation translation vector based on the maximum absolute value in a processing window to generate the deformation parameter; Comparing the gesture fluctuation index with a preset correction grid threshold, triggering high-density self-adaptive correction grid mapping when the gesture fluctuation index is judged to be larger than or equal to the correction grid threshold, and reducing the pixel span of the grid side length according to the gesture fluctuation index; and acquiring the coordinate correction coefficients of the grid nodes based on the perspective transformation function, the three-dimensional space rotation matrix and the deformation parameters, and performing space resampling on the original image by using a bilinear interpolation algorithm to generate the corrected image data.
  5. 5. The unmanned aerial vehicle edge calculation landslide segmentation modeling method based on dynamic gesture correction according to claim 1, wherein preprocessing the corrected image data, extracting and screening edge lines to determine a preliminary contour range specifically comprises: Denoising and perspective correction are carried out on the corrected image data by applying a two-dimensional median filter and a distortion removal algorithm based on a polynomial distortion model to obtain preprocessed image data; Generating a preliminary edge distribution map in the preprocessed image data by adopting an edge detection algorithm and combining non-maximum suppression and a double-threshold detection mechanism; performing spatial domain weighted smoothing on a coordinate sequence of continuous edge pixels in the preliminary edge distribution map by using a one-dimensional Gaussian kernel function to generate a smooth edge line serving as the edge line; calculating the ratio of the connection length of adjacent edge points to the length of the total edge points of the edge lines as a continuity index, and screening the edge lines reaching a preset continuity threshold according to the continuity index to be classified into a continuous edge set; extracting the edge lines with geometric features in the continuous edge set by using Hough transformation, and determining the primary contour range by using a morphological expansion operator and a corrosion operator to perform closure processing.
  6. 6. The unmanned aerial vehicle edge calculation landslide segmentation modeling method based on dynamic gesture correction according to claim 1, wherein determining that the power-calculation scheduling branch is in a stable state or a fluctuation state according to the gesture fluctuation index according to the linkage mapping rule, performing position update in the stable state, and triggering a classification mechanism in the preliminary contour range in the fluctuation state to determine a disaster core area specifically comprises: When the stationary state is judged, extracting a history core area boundary determined by the previous frame as a tracking feature, calculating pixel displacement increment between adjacent frames by using a sparse optical flow algorithm, and overlapping the pixel displacement increment on coordinates of the history core area boundary to execute the position update; When the fluctuation state is judged, constructing a gray level co-occurrence matrix quantized pixel contrast characteristic value with a designated step length and a designated direction in the primary contour range; And extracting geometric form features of the preliminary contour range, splicing the pixel contrast feature values and the geometric form features to generate high-dimensional feature vectors, inputting the high-dimensional feature vectors into a support vector machine adopting a radial basis function to execute nonlinear classification mapping as the classification mechanism, and removing an interference area to determine the disaster core area.
  7. 7. The unmanned aerial vehicle edge calculation landslide segmentation modeling method based on dynamic gesture correction according to claim 2, wherein the extracting the visual translation amount of the multi-frame correction image data in the disaster core area, and triggering closed loop calibration when the vector difference value between the visual translation amount and the physical deformation translation vector shows an increasing trend specifically comprises: capturing motion vectors of pixels of a plurality of frames of corrected image data in the disaster core area by using an optical flow algorithm, removing abnormal vectors in the motion vectors by using a random sampling consistency algorithm, and then solving an average value to extract the visual translation quantity; performing vector subtraction in a feature alignment space to obtain the vector difference between the visual translation amount and the physical deformation translation vector; And continuously executing time integration on the vector difference value within a preset observation period time window to monitor a modular length evolution track of the accumulated drift error, and judging that the accumulated drift error is generated to trigger the closed loop calibration when the modular length evolution track of the accumulated drift error is detected to show a continuous unidirectional increasing trend as the increasing trend.
  8. 8. The unmanned aerial vehicle edge calculation landslide segmentation modeling method based on dynamic gesture correction of claim 7, wherein the triggering closed-loop calibration to reconstruct the coordinate correction coefficients, outputting a dynamic boundary description specifically comprises: Dynamically amplifying diagonal elements of the measurement noise covariance matrix by using the vector difference as a feedback gain variable through an adaptive adjustment function; The updated measurement noise covariance matrix is acted on the fusion processing at the next moment, and the coordinate correction coefficient is synchronously reconstructed; after reconstructing the coordinate correction coefficient, extracting an edge coordinate point set of the disaster core area of the current frame, and introducing a density-based spatial clustering algorithm to remove discrete noise points which do not meet the density reachable condition; And integrating the space-time sequence features of the edge coordinate point set after eliminating the discrete noise points, and outputting the dynamic boundary description.
  9. 9. The unmanned aerial vehicle edge computing landslide segmentation modeling method based on dynamic attitude correction according to claim 1, wherein extracting geometric element data described by the dynamic boundary, evaluating the stability of a local boundary, and dividing the geometric element data into a stable data set and an unstable data subset according to the stability specifically comprises: extracting a geometric element coordinate set from the dynamic boundary description as geometric element data, and performing track smoothing on an abscissa sequence parameterized according to the boundary perimeter by adopting one-dimensional Kalman filtering; Performing segment sampling on the abscissa sequence after track smoothing, fitting a reference straight line by adopting a local least square method, and calculating the vertical Euclidean distance from a sampling point to the reference straight line to obtain the geometric standard deviation of the local boundary as the stability; and dividing the boundary line segment with the geometric standard deviation lower than a preset stability threshold into the stable data set, and dividing the boundary line segment with the geometric standard deviation greater than or equal to the preset stability threshold into the unstable data subset.
  10. 10. The unmanned aerial vehicle edge computing landslide segmentation modeling method based on dynamic posture correction of claim 9, wherein reorganizing the unstable data subset and the stable data set to generate landslide segmentation results specifically comprises: performing mean value filtering and noise reduction on the unstable data subsets by adopting a time window moving average method, and performing endpoint splicing, fusion and recombination on the unstable data subsets subjected to mean value filtering and noise reduction and the stable data sets by adopting a B spline curve interpolation algorithm to form closed geometric shape point sets; eliminating redundant boundary points with the geometric shape point concentration distance connecting lines smaller than a preset distance tolerance threshold by using a curve simplification algorithm to obtain key topology nodes; And connecting the key topological nodes according to a topological sequence to perform polygon fitting, and generating landslide disaster peripheral polygons as landslide segmentation results by combining a clustering algorithm.

Description

Unmanned plane edge calculation landslide segmentation modeling method based on dynamic attitude correction Technical Field The invention relates to the technical field of image processing, in particular to an unmanned aerial vehicle edge calculation landslide segmentation modeling method based on dynamic gesture correction. Background In the disaster monitoring field, unmanned aerial vehicle edge computing technology becomes an important means for acquiring geological disaster site information such as landslide and the like due to the flexible and efficient characteristics. The visual sensor carried by the unmanned aerial vehicle is used for acquiring disaster site images, and real-time landslide area segmentation is carried out at the edge end, so that basic data can be provided for subsequent disaster evaluation and monitoring. Accurate landslide disaster boundary information with geometric continuity is obtained, and is a precondition for improving disaster monitoring quality. In the prior art, when an unmanned aerial vehicle landslide image under a complex flight environment is processed, single-source visual information is mainly relied on for processing, and physical flight attitude data of the unmanned aerial vehicle cannot be fully combined. When unmanned aerial vehicle receives the air current influence to produce the high frequency shake, the image of shooting can take place perspective deformation and distortion, owing to lack the closed loop calibration mechanism between physical translation and the vision translation, leads to splitting the spatial accuracy decline of boundary. Meanwhile, the computing force resources of the edge computing equipment are limited, the traditional method generally adopts a fixed algorithm model to carry out continuous computation, the linkage mapping relation between computing force scheduling and flight attitude fluctuation is not established, and workflow switching can not be carried out under the stable flight and severe fluctuation state, so that the equipment is difficult to realize the on-demand distribution of computing force, an interference area is easily introduced when a machine body shakes, and the extraction precision of a disaster core area is reduced. In addition, the existing boundary fitting means lacks of evaluating the stability of the local boundary, and cannot implement differentiated reorganization and noise reduction smoothing processing on unstable boundary segments, so that the generated landslide segmentation result retains more saw-tooth distortion, and a coherent polygonal contour with boundary closure degree and low geometric redundancy is difficult to form. Therefore, how to overcome the influence of unmanned aerial vehicle dynamic flight attitude on image deformation, and combining a computational power scheduling mechanism and a boundary difference smoothing means to realize accurate and stable landslide boundary segmentation in a complex environment is a problem to be solved in the field. Disclosure of Invention Aiming at the defects of the prior art, the invention provides an unmanned aerial vehicle edge calculation landslide segmentation modeling method based on dynamic gesture correction, which solves the problems that unmanned aerial vehicle images generate deformation distortion due to high-frequency jitter, edge calculation equipment lacks dynamic adaptability in calculation force distribution in a complex flight state, and local boundary distortion exists in landslide segmentation results. The invention provides an unmanned aerial vehicle edge calculation landslide segmentation modeling method based on dynamic attitude correction, which adopts the following technical scheme: the unmanned aerial vehicle edge calculation landslide segmentation modeling method based on dynamic attitude correction comprises the following steps: acquiring an original image acquired by an unmanned aerial vehicle, synchronously acquiring multidimensional gesture data, and carrying out fusion processing on the multidimensional gesture data based on a dynamic fusion model to acquire a complete gesture data set containing gesture description parameters; extracting dynamic characteristics of the complete gesture data set to calculate gesture fluctuation indexes, and establishing a linkage mapping rule of the gesture fluctuation indexes and a computational power scheduling branch; According to the gesture description parameters, resolving deformation parameters comprising physical deformation translation vectors, acquiring coordinate correction coefficients based on the deformation parameters and combining the gesture fluctuation indexes, and correcting the original image by using the coordinate correction coefficients to generate corrected image data; preprocessing the corrected image data, extracting and screening edge lines to determine a preliminary contour range; Judging that the power calculation scheduling branch is in a stable state or a fluctu