40 results about "Homography" patented technology

The invention discloses a building outer wall falling risk detection method and system based on deep learning, and the method comprises the steps: obtaining an original surface image, extracting vertical face geometric features, constructing a homography matrix based on vanishing points, carrying out the geometric correction, and reconstructing an orthographic projection image; performing enhancement processing on the orthographic projection image, extracting an effective detection area and intercepting a detection image; inputting the detection image into a deep learning segmentation model into which edge weight constraint is introduced, and performing semantic segmentation, binaryzation and optimization to obtain a falling region; a mapping relation is established according to the actual physical size of the building outer wall, and pixel information of the falling area is converted into multi-dimensional parameters such as the actual physical area, the elevation, the width and the height; and combining the defect intensity input quantification risk assessment model to calculate a comprehensive risk index, determining a risk level and outputting a result. Automatic identification and quantitative evaluation can be realized, distortion and background interference are eliminated, accurate conversion from pixels to a multi-dimensional physical space is realized, and a scientific basis is provided for safety investigation.

The invention discloses a method for optimizing micro-crack segmentation based on deep learning and super-resolution reconstruction, and the method comprises the steps: cutting an image in real time, obtaining an image block which takes a component as a target main body, and synchronously recording a homography matrix for geometric mapping; selecting an amplification strategy to improve the resolution, and recording a scale mapping relation; inputting the enhanced image block into a double-flow network; adaptive fusion and reconstruction are carried out on the two branch features, and a high-resolution texture image is output; generating a geometrically corrected ortho-image, fusing the geometrically corrected ortho-image with original illumination information, and outputting a corrected image with a known pixel size; identifying cracks, spalling and honeycomb diseases in parallel; generating a unified defect confidence map; calculating real geometric parameters of the BIM in a BIM global coordinate system through coordinate back projection; and generating quantitative defect reports and maintenance suggestions. The method has the advantage that seamless connection between the detection result and the BIM global coordinates is realized.

The invention relates to the technical field of industrial vision and intelligent monitoring, and discloses a loom control system operation state monitoring and fault diagnosis method based on machine vision, which comprises the following steps: acquiring a video stream in a loom shed area and constructing a two-dimensional space-time slice tensor; performing global motion compensation processing on the space-time slice tensor by using a homography transformation matrix, mapping a compensated dynamic texture feature sequence to a three-dimensional phase space by using a time delay embedding algorithm, and reconstructing a closed phase space trajectory representing periodic operation logic of the loom; the discrete Frechet distance between the phase space trajectory of the current operation cycle and the preset reference trajectory is calculated, and a control instruction is generated. The health degree of the sequential logic of the system is directly quantified on the premise that specific components are not recognized by using the invariant characteristic of the phase space manifold topology; the technical problems that small phase lag is difficult to perceive and nonlinear faults cannot be early warned in a strong noise environment are solved.

The invention provides a three-dimensional reconstruction method based on guided filtering and Mama geometric feature fusion in the technical field of three-dimensional reconstruction, and the method comprises the steps: S1, collecting a multi-view RGB image, carrying out the calibration of a camera, obtaining a projection matrix and an internal reference matrix, and obtaining a reference image; s2, extracting image features from each RGB image; s3, sampling on the reference image to obtain a depth hypothesis plane, performing homography transformation on each image feature, and aggregating with the depth hypothesis plane to obtain a cost body; s4, sequentially performing filtering, regularization and exponential normalization on the cost body to obtain a probability body, and constructing a depth map and a confidence map based on the probability body; s5, performing frequency domain filtering and up-sampling on the depth map and the confidence map, and then performing feature fusion to obtain a visual angle depth; and S6, executing three-dimensional reconstruction operation based on the visual angle depth, the projection matrix and the internal reference matrix. The method has the advantage that the precision and robustness of three-dimensional reconstruction are greatly improved.

The invention discloses a cross-view multi-target collaborative perception and global tracking method. The method comprises the following steps: receiving original video streams of different views in parallel; carrying out feature extraction and content perception feature recombination processing; establishing a homography transformation matrix representing the mapping relation between the visual angles; performing forward geometric mapping by using the homography transformation matrix, and performing reverse mapping, secondary confirmation and state updating; a joint cost model based on space-time-appearance fusion is constructed, the potential association degree of unassociated residual targets is quantified, and missing targets are recalled; performing duplicate removal and optimization on the targets in the association set; and training and optimizing the network. According to the embodiment of the invention, the method achieves the capturing of a pixel-level target, and improves the texture and edge expression capability of a tiny target in a deep feature map. Efficient fusion of multi-dimensional spatio-temporal information among visual angles is realized; higher robustness and tracking continuity are achieved, and the integrity and fineness of a final trajectory result are improved; and the recall is ensured, and the generation of redundant tracks is avoided at the same time.

The invention provides an image registration method and an image registration device, which are applied to the technical field of image processing. Comprising the following steps: acquiring a target reference image and a first target to-be-registered image of a to-be-detected area containing FOD; constructing an optimal affine transformation model according to the target reference image and the N preset pitch angles, and further performing linear transformation on the first target to-be-registered image to obtain a second target to-be-registered image; according to a Gaussian difference pyramid of a second target to-be-registered image, screening an extreme point of which the principal curvature does not exceed a preset principal curvature threshold as a first feature point, and further screening a first feature point of which the re-projection error and the sampling probability meet a preset condition as a second feature point; and solving a homogeneous linear equation set of the pixel point pairs matched with the second feature points by adopting an SVD method to obtain a transformation homography matrix, and then registering the first target to-be-registered image to obtain a target image. The contradictory closed loop of low efficiency-insufficient precision is broken through, and the aviation operation safety is guaranteed.

For a cross-modal target prompt matching task, the invention provides a training-free multi-stage matching method based on local feature points and global spatial transformation so as to realize robust positioning of the same target in different modal images. According to the method, firstly, feature points and descriptors of a target prompt image and a to-be-matched image are extracted by utilizing a pre-trained SuperPoint, then, MINIMAGlue finely adjusted on cross-modal data is preferentially adopted to perform feature point matching, and when matching fails, LightGlue is switched to serve as remediation; in the matching process, an optimal angle is selected through a rotation enumeration strategy to enhance matching robustness; in the coarse positioning stage, calculating a minimum enclosing rectangle according to matching points in the ROI, and if the matching points are insufficient, estimating a homography matrix by adopting RANSAC and carrying out geometric correction on a rectangular frame; in the fine positioning stage, the central point of the coarse positioning result is used as a prompt to input MobileSAM for target segmentation, and a final rectangular frame is corrected through area consistency constraint. Through integration and fusion of the multi-stage process, accurate matching and stable positioning of cross-modal target prompt are realized under a training-free condition, and the generalization ability and practicability of the method are improved.

The application provides a feature matching method, system and device based on multi-geometry collaborative learning and a storage medium, and relates to the technical field of robot positioning and navigation. The steps include: acquiring an image pair, extracting two groups of sparse local features in the image pair and inputting the local features into a multi-geometry collaborative network model; enhancing the local features through an attention mechanism and iteratively updating the sparse network through a multi-layer perception and / or linear projection; generating and screening reliable matching points through affine transformation estimation, position coding and key point neighborhood expansion; constraining the matching process through pose information and adaptively adjusting the matching strategy for matching prediction; combining real matching information and homography geometry information for loss calculation to train the multi-geometry network model; and repeating the steps until a preset iteration number is reached. Through collaborative optimization of affine geometry, epipolar geometry and homography geometry, the application realizes mutual promotion of local feature discriminability and geometric consistency, and can achieve a better balance between speed and accuracy in feature matching.

The invention relates to the technical field of defect detection, in particular to a railway contact line surface defect detection method and system. The method comprises the steps of obtaining a sequence image; calculating a space-time matching weight based on the tracking duration and the normal distance of the feature point pair, and obtaining a global homography matrix through a weighted random sampling consensus algorithm for preliminary correction; calculating a local affine transformation matrix and a distortion factor for the image segment, and splicing to generate a two-dimensional expansion graph; and projecting the candidate defect area segmented on the expanded image back to the original image, calculating a gradient direction entropy ratio, and judging a real defect based on a self-adaptive threshold determined by a distortion factor. According to the scheme, non-rigid deformation of the contact line can be compensated, geometric distortion and real defects can be distinguished in a self-adaptive mode, and the detection accuracy is improved.

This application discloses a real-time image stitching method and apparatus. The method includes: receiving images uploaded to the cloud by a camera device in real time; extracting feature points from the images and matching feature points between the current frame and the previous frame; calculating the ratio of the homography matrix score to the sum of the homography matrix score and the fundamental matrix score to determine whether to use the homography matrix or the fundamental matrix to recover the image's pose parameters and map points, and initializing the map; estimating the pose of the current frame and optimizing the pose of the current frame using candidate frames; converting the coordinates of the map points to two-dimensional coordinates and performing plane fitting; converting the coordinates of the image corner points in the camera coordinate system to the object coordinate system, calculating perspective transformation parameters and performing geometric transformation on the image; and performing Gaussian pyramid fusion on the generated tiles. This method solves the problems of low efficiency in scene reconstruction using motion reconstruction algorithms and the inability to process real-time photogrammetric data.

The invention relates to the technical field of three-dimensional scanning, in particular to a point cloud enhancement method and device based on binocular reconstruction, equipment and a storage medium. The method comprises the following steps: acquiring sparse point cloud; projecting the sparse point cloud to a preset left pixel plane and a preset right pixel plane to obtain a left pixel coordinate and a right pixel coordinate; determining a bounding box area according to the distribution relation of the left pixel coordinates, and determining a matching point pair set about the left pixel coordinates and the corresponding right pixel coordinates by traversing pixel points in the bounding box area; calculating a local homography matrix and an average depth value according to the matching point pair set, generating a matching point of a right pixel coordinate by using the homography matrix, performing correction through epipolar constraint, and performing three-dimensional reconstruction on the corrected matching point pair to obtain a reconstruction point; and verifying the reconstruction points according to the average depth value, and adding the reconstruction points meeting verification conditions to the global point cloud. According to the method, part of discrete and sparse point clouds can be effectively enhanced.

The invention relates to the technical field of industrial automation, in particular to a strain gauge front and back face recognition and mounting method based on visual recognition, which comprises the steps of dual polarization and speckle collection, radiation correction and geometric correction, Stokes vector calculation and target grid generation, grid-level Mueller inversion and phase statistics. Constructing an object-level evidence and forming attitude initial estimation, executing two-out-of-three consistent judgment according to a threshold and re-checking by using a tail end polarization image, and performing circle statistical fusion on the hand attitude and the attitude initial estimation to obtain a final attitude; and projecting the mounting reference to a station plane through the homography matrix, synthesizing a mounting pose and issuing an instruction. According to the method, stable judgment is kept in light reflection, weak texture and batch change scenes, the error pasting rate is reduced, and process tracing and online calibration are supported.

The application discloses a noise elimination and transformation model robust estimation method in multi-modal image registration, comprising: obtaining initial matching point pairs, establishing a homography transformation model to be estimated and residual error measurement; performing vector field consistency calculation on the initial matching point pairs, and constructing an edge expansion consensus set; performing initial fitting on the homography transformation model on the edge expansion consensus set to obtain an initial homography transformation model; estimating noise scale and initializing residual error threshold based on residual error statistics induced by the initial homography transformation model; regarding the residual error threshold and the homography transformation model parameters as same-order decision quantities, constructing a joint optimization framework of threshold-model cooperation; adaptively constructing a joint search domain according to regularity statistics of residual error distribution; constructing a threshold-coupled cost function under a mixed statistical modeling framework of inliers / outliers; performing collaborative iterative updating on the homography transformation model and the residual error threshold by adopting a group optimization strategy, and outputting the homography transformation model and a refined inlier set.

The invention discloses an industrial scene anomaly detection method based on multi-dimensional feature decoupling and a double-track state machine, and the method comprises the steps: taking an industrial field monitoring video stream as data input, firstly carrying out the spatial correction of the video stream through a feature anchor point-based dynamic registration algorithm and a homography matrix, and eliminating background jitter; secondly, the correction frames are input into a heterogeneous channel in parallel, and an initial foreground is generated through adaptive Gaussian mixture and HSV combined filtering; thirdly, reconstructing a target by using morphological filtering, inputting a double-track time sequence module, and shunting the double-track time sequence module to an existing track or a persistent track for judgment; and finally, integrating a time sequence accumulation result of each track, and outputting an alarm signal containing an abnormal type, coordinates and confidence. The method is suitable for complex scenes with numerous interferences, effectively solves the problems of jitter false alarm and mechanical retention missing detection, and improves the detection accuracy.

This application relates to the field of image processing technology for fly traps, providing an adaptive stitching method and system for fly trap images. The method includes: extracting feature vectors from multiple images using the SIFT algorithm and a ResNet network respectively, and weighting and fusing the two features using adaptive weight coefficients 'a' generated by the ResNet network to obtain a fused feature vector; based on the fused feature vector, using an improved random sampling consensus algorithm to solve the homography matrix between the two images, performing non-uniform random sampling based on similarity during the sampling stage, and using the top K% of the most similar inliers to form a refined inlier set for calculation during the calculation stage; and completing image stitching based on the homography matrix. This application solves the technical problems of difficult feature matching, stitching misalignment, and obvious fusion traces caused by the small size, dense distribution, and high similarity of insect targets in fly trap images, thus improving the robustness, accuracy, and efficiency of image stitching.

The invention discloses a foundation pit supporting structure surface defect detection system and method based on a hierarchical graph model, and the method comprises the steps: firstly carrying out the construction and data collection of a detection system, then extracting a supporting structure surface region from an obtained original data image through a semantic segmentation model, and removing a non-target background; detecting a defect mask in the surface area of the supporting structure after the non-target background is removed by using an improved YOLOv11-seg instance segmentation model, and generating an independent mask; secondly, geometric index calculation of the area, the length and the width is carried out on each obtained defect mask, a pixel unit is converted into an actual physical unit in combination with a homography transformation matrix H, and it is ensured that a detection result can be directly used for structural safety evaluation; and then evaluation and VLM expert judgment are carried out, and finally a diagnosis report is generated and a result is visualized. According to the automatic crack analysis system, macroscopic evaluation and microscopic analysis can be integrated, computing resources can be intelligently managed, and environmental noise can be effectively resisted.

The invention provides a multi-view target tracking method which can be applied to the technical field of artificial intelligence. The multi-view target tracking method comprises the following steps: extracting feature points from a synchronous image, and obtaining a time sequence homography matrix of the synchronous image according to the feature points; matching the reference points of the multiple targets of the at least two view angles to obtain matching point pairs; obtaining a cross-view homography matrix according to the matching point pairs; further obtaining a comprehensive transformation matrix of the plurality of targets between at least two visual angles; establishing tracks of the multiple targets according to the reference points of the multiple targets at different time; projecting the tracks of the at least two visual angles into one visual angle by using the comprehensive homography matrix; and matching the tracks and the projections of the plurality of targets according to the similarity of the tracks and the projections of the plurality of targets so as to perform target tracking. The invention also provides a multi-view target tracking device, equipment, a storage medium and a program product. According to the embodiment of the invention, the robustness of target association in a complex scene can be improved.

The invention relates to a building three-dimensional reconstruction method based on a multi-coplanar geometry and graph neural network. The building three-dimensional reconstruction method comprises the following steps: preprocessing an aerial image and a ground image and obtaining an initial point matching set; iteratively extracting a plurality of groups of local homography matrixes between the image pairs; scaling the aerial image and the ground image and constructing multi-scale pyramid representation; a deep matching network is adopted for matching; matching results are gathered and then projected back to an original image coordinate system through inverse transformation of a homography matrix to form a candidate cross-platform corresponding point-line set; and replacing the matched input of the reconstruction process with the verified cross-platform control point set to output a final building three-dimensional model. According to the method, a sufficient number of cross-platform control points which are uniformly distributed and geometrically consistent can be obtained from cross-platform image pairs with large view angle difference, scale change and local shielding, so that matching input of traditional three-dimensional reconstruction is replaced or reinforced, and high-precision and high-integrity three-dimensional reconstruction and detail recovery of an urban building scene are realized.

The application relates to a building three-dimensional reconstruction method based on multi-coplanar geometry and a graph neural network, which comprises the following steps: pre-processing aerial images and ground images and obtaining an initial point matching set; iteratively extracting multiple groups of local homography matrices between the image pairs; scaling the aerial images and the ground images and constructing a multi-scale pyramid representation; performing matching by using a deep matching network; after the matching results are converged, the inverse transformation of the homography matrix is used to project back to the original image coordinate system to form a candidate cross-platform corresponding point line set; and verified cross-platform control point set is used to replace the matching input of the reconstruction process to output a final building three-dimensional model. The application can obtain sufficient, uniform and geometrically consistent cross-platform control points in cross-platform image pairs with large angle differences, scale changes and local occlusions, thereby replacing or reinforcing the matching input of traditional three-dimensional reconstruction, and realizing high-precision, high-completeness three-dimensional reconstruction and detail recovery of a city building scene.

The application provides an image registration method and an image registration device, and applies to the technical field of image processing. The method comprises the following steps: obtaining a target reference image and a first target image to be registered of a detection area containing FOD; constructing an optimal affine transformation model according to the target reference image and N preset pitch angles, and then performing linear transformation on the first target image to be registered to obtain a second target image to be registered; screening extreme points with a principal curvature not exceeding a preset principal curvature threshold as first feature points according to a Gaussian difference pyramid of the second target image to be registered, and then screening the first feature points satisfying a preset condition in re-projection error and sampling probability as second feature points; solving a homogeneous linear equation set of pixel pairs matched with the second feature points by using an SVD method to obtain a transformation homography matrix, and then performing registration on the first target image to be registered to obtain a target image. The method breaks through the contradiction closed loop of "low efficiency and insufficient accuracy", and guarantees the safety of aviation operation.

A high-generalization landing point localization method based on small sample data from extraterrestrial topographic images is proposed. The method acquires the DOM (Domain Image of Landing Area) of the spacecraft landing zone and the original descent sequence images. The DOM is segmented into small blocks, denoted as sub-base maps. Adaptive frame extraction is performed on the descent sequence images to calculate the homography matrix between adjacent frames and recover the relative pose of the cameras. The estimated altitude of each frame is obtained, and the resolution of each frame in the descent sequence is calculated. The descent image closest to the physical resolution of the DOM is selected as the descent image keyframe. Exhaustive matching is performed between the descent image keyframe and all sub-base maps to select the optimal sub-base map. Matching point pairs between adjacent descent sequence images are obtained. Combining the matching point pairs between the descent image keyframe and the optimal sub-base map, the coarse latitude and longitude coordinates of the landing point are obtained. The optimal sub-base map is added to the descent sequence for clustering adjustment to obtain the fine latitude and longitude coordinates of the landing point.

The application discloses a hybrid homography transform rolling shutter single image correction system and method, and a computer storage medium. The method comprises the following steps: dividing an image to be processed into K blocks, and assigning a weight to each motion base of a homography matrix of each block; processing the K blocks respectively to obtain an initial homography transform flow field corresponding to each of the K blocks; defining a Gaussian weight along a center line of the K blocks by using a Gaussian smoothing feature to obtain K Gaussian weight maps; performing a dot product on the initial homography transform flow field corresponding to each of the K blocks and the K Gaussian weight maps to obtain a final homography transform flow field of the image to be processed; and mapping the image to be processed based on the final homography transform flow field of the image to be processed to obtain a corrected image. Compared with a traditional single homography transform, the hybrid homography transform has a higher spatial degree of freedom and can be used for rolling shutter correction.

The invention discloses a wide parallax image splicing method based on epipolar constraint and structure maintenance. Firstly, feature extraction and matching are carried out on image pairs, initial feature point pairs and feature line pairs are obtained, and mismatched feature point pairs are removed; thirdly, candidate feature point pairs are generated through the feature line pairs, coplanarity screening is carried out on the candidate feature point pairs through epipolar constraint, and a geometric consistent feature matching set is constructed; secondly, estimating a global homography transformation matrix and performing pre-alignment by utilizing the geometric consistency feature matching set; uniformly dividing a to-be-spliced image into a plurality of grid regions, optimizing the grid regions, and calculating a local homography transformation matrix of the grid regions; finally, estimating a global similarity transformation matrix and respectively combining the global similarity transformation matrix with the local homography transformation matrix to obtain a transformation matrix of the grid region; and projecting a to-be-spliced image by using the transformation matrix of the grid region, and fusing to generate a panoramic image. Through combination of global and local transformation, the splicing precision of the wide parallax image is improved.

The invention belongs to the technical field of non-contact three-dimensional measurement, discloses a galvanometer type line laser three-dimensional reconstruction method based on a double BP neural network, and aims to solve the problems of complex calibration, low laser line extraction precision and large coordinate calculation error in the prior art. The method comprises the following steps: firstly, fixing poses of a galvanometer and line laser, calibrating internal reference of a camera and establishing a coordinate system; a data set is acquired through Z-axis translation of a calibration plate and multi-angle scanning of a galvanometer, laser rays are searched by adopting a depth-first algorithm, and center pixel coordinates of stripes are extracted from each column in a Weibull distribution manner; a precise three-dimensional coordinate data set is obtained by combining a homography matrix and light plane fitting, a double BP neural network is constructed, and XY coordinates and Z-axis coordinates are sequentially trained and output; during actual measurement, laser line characteristic data and a galvanometer angle of a measured object are input, and three-dimensional reconstruction is completed. The calibration process is simplified, the laser line extraction and coordinate calculation precision is improved, the cost is low, the adaptability is high, and the method is suitable for industrial detection, reverse engineering and other scenes.

The invention discloses a noise elimination and transformation model robust estimation method in multi-modal image registration, and the method comprises the steps: obtaining an initial matching point pair, and building a to-be-estimated homography transformation model and a residual measurement; vector field consistency calculation is carried out on the initial matching point pairs, and an edge expansion consensus set is constructed; performing initial fitting on the homography transformation model on the edge expansion consensus set to obtain an initial homography transformation model; estimating a noise scale based on residual statistics induced by the initial homography transformation model and initializing a residual threshold; regarding a residual threshold and a homography transformation model parameter as same-order decision quantities, and constructing a threshold-model collaborative joint optimization framework; a joint search domain is constructed in a self-adaptive mode according to regularity statistics of residual error distribution; constructing a threshold coupling cost function under an inner point / outer point mixed statistical modeling framework; and carrying out collaborative iteration updating on the homography transformation model and the residual threshold by adopting a group optimization strategy, and outputting the homography transformation model and a refining inner point set.

The invention provides a visual measurement static feature point tracking and error suppression method based on a homography mapping chain, and the method comprises the steps: shooting a video of a to-be-measured target and a video of a static feature point at the same time, carrying out the enhancement and filtering of an image, selecting an ROI to recognize the to-be-measured target and the static feature point, and obtaining the real-time displacement data of the to-be-measured target and the static feature point through a feature point matching algorithm; feature points which cannot be stably matched or have large matching errors are screened and marked, and the optimal estimation of the feature points in the image is solved by adopting a homography mapping chain algorithm; correcting the coordinates of the feature points through the optimal estimation obtained through calculation; the static feature point displacement, namely the measurement error, is calculated through a weighted distance neighborhood algorithm; and designing an adaptive error suppressor through the measurement error and the displacement of the to-be-measured target, and calculating to obtain the denoised world displacement of the to-be-measured target. According to the method provided by the invention, the problem that the static feature points cannot be matched or are wrongly matched in a long-time tracking process is solved, and the environmental adaptability of the visual measurement error suppression method is improved.

The embodiment of the invention provides a panoramic image interpolation method and device. Comprising the following steps: determining a point I0 on a panoramic longitude and latitude graph I; mapping to a fisheye image according to a conversion model, and taking four adjacent integer coordinate adjacent points; a unit sphere S is determined according to the panoramic longitude and latitude graph I, the panoramic longitude and latitude graph I is pasted to the unit sphere, plane mapping is carried out by taking the panoramic longitude and latitude graph I as the center, a projection plane P is obtained, calculation is carried out by using a least square method, and a homography matrix is obtained; and completing interpolation according to the homography matrix. Through the steps of mapping, projection, matrix solving and interpolation, the edge sawtooth problem is optimized, meanwhile, the accuracy and universality of the interpolation process are guaranteed, and the method is suitable for various panoramic unfolding scenes based on fisheye images.

This disclosure provides a training method and apparatus for a homography flow estimation model, a method and apparatus for homography flow estimation based on monocular images, a computer-readable storage medium, and an electronic device. The training method includes: extracting features from a first sample image and a second sample image in the same image sequence using a model to be trained, obtaining feature data; performing prediction using the model to be trained based on the feature data, obtaining predicted homography flow data; determining a homography flow loss value based on the predicted homography flow data and a preset homography flow loss function; and adjusting the parameters of the model to be trained based on the homography flow loss value to obtain a homography flow estimation model. This disclosure enables the training of a homography flow estimation model, allowing for the prediction of ground-based homography flow data from image sequences captured by a monocular camera, significantly improving the efficiency and accuracy of homography flow data prediction.

The application provides a three-dimensional reconstruction method based on guided filtering and Mamba geometric feature fusion in the technical field of three-dimensional reconstruction, and comprises the following steps: S1, collecting multi-view RGB images, calibrating the camera to obtain a projection matrix and an intrinsic parameter matrix, and obtaining a reference image; S2, extracting image features from each RGB image; S3, sampling a depth hypothesis plane on the reference image, performing homography transformation on each image feature, and aggregating the depth hypothesis plane to obtain a cost volume; S4, sequentially performing filtering, regularization and exponential normalization on the cost volume to obtain a probability volume, constructing a depth map and a confidence map based on the probability volume; S5, performing frequency domain filtering and upsampling on the depth map and the confidence map, and then performing feature fusion to obtain a view depth; and S6, performing a three-dimensional reconstruction operation based on the view depth, the projection matrix and the intrinsic parameter matrix. The application has the advantages that the accuracy and the robustness of three-dimensional reconstruction are greatly improved.

The application discloses a multi-view three-dimensional reconstruction method and system based on GRU and 3DCNN, and the method comprises the following steps: performing feature extraction on images of a target object under different visual angles to obtain feature maps; performing differentiable homography transformation on each feature map under different depth assumptions; dividing a source feature body and a reference feature body into multiple groups according to channels, calculating group similarity and aggregating a matching cost body; performing cost body regularization operation and calculating a depth expectation value; upsampling and normalizing a low-resolution depth map; obtaining an updated depth map according to the matching cost body and constructing a dynamic cost body; constructing a matching cost body to obtain a depth map with the same size as the original resolution; and finally completing multi-view three-dimensional reconstruction according to a corresponding loss function. The embodiment of the application can improve the high-resolution depth estimation accuracy and realize more complete three-dimensional reconstruction, and can be widely applied to the field of computer technology.