57 results about "Homography (computer vision)" patented technology

The invention relates to the field of computer vision and three-dimensional reconstruction, and particularly discloses a multi-view three-dimensional reconstruction method based on multi-scale feature fusion, which adopts a multi-scale feature fusion network architecture and comprises a feature extraction module, a cost body construction module, a cost body regularization network and a deep regression network. The method specifically comprises the following steps: constructing a multi-view three-dimensional network architecture based on multi-scale feature fusion; the method comprises the following steps: setting a feature extraction module, combining an FPN structure with a bidirectional feature fusion strategy, establishing a homography transformation and cost body construction module, generating a cost body, deploying a cost body regularization processing unit, carrying out global-local feature fusion by adopting a 3D UNet architecture, and executing a depth inference and refinement process. And optimizing a depth regression result through multi-stage depth hypothesis and a confidence weighting mechanism. And performing a three-dimensional reconstruction experiment and result analysis, implementing depth map fusion and three-dimensional reconstruction, and generating a dense three-dimensional point cloud model. According to the invention, the precision and integrity of three-dimensional reconstruction are effectively improved.

The invention discloses a multi-target tracking method combining camera motion compensation and pseudo depth estimation, belongs to the field of computer vision, and is suitable for a complex automatic driving road environment. The method comprises the following steps: constructing a training set and a test set; detecting the image by using a deep learning detector and extracting features; a Kalman filter is adopted to correct a motion modeling state vector, and the target position and size prediction precision is improved; solving a homography matrix through feature point matching, performing global camera motion compensation, and reducing camera jitter and displacement interference; target pseudo depth information is calculated, hierarchical cascade matching is carried out, and association performance in dense and shielding scenes is optimized; a three-level cascade strategy is adopted to complete high confidence degree, low confidence degree and residual target matching in sequence; and finally, outputting a tracking result with a detection frame and identity information to obtain a trained model. According to the invention, accurate detection and stable tracking of multi-category targets can be realized in a complex environment, and identity switching is effectively reduced.

The invention relates to the technical field of computer vision, discloses a personnel positioning method, device and equipment based on detection frame correction, and is used for solving the technical problem of low personnel positioning precision caused by detection frame errors. The personnel positioning method based on detection frame correction comprises the steps that whether an initial detection frame meets preset positioning reference constraint conditions or not is determined through human body key points located in the initial detection frame, and the positioning reference constraint conditions at least include that both a left frame line and a right frame line of the initial detection frame do not have single-side deviation; if the left frame line or the right frame line has single-side offset, a shoulder joint point on the same side as the offset side is selected as a target key point, a target offset value is determined in combination with the distance between the corresponding frame lines in the initial detection frame, and the detection frame is moved and corrected, so that the corrected target detection frame meets a positioning reference constraint condition; and positioning the pixel coordinates of the midpoint of the bottom frame line of the target detection frame and a preset target homography matrix.

The invention discloses an automatic cattle checking method and system based on two-dimensional panoramic vision, and belongs to the technical field of computer vision and intelligent breeding, and the method comprises the following steps: reasonably arranging a plurality of fixed cameras in a cattle farm area, obtaining the internal reference and distortion coefficient of each camera through employing a camera calibration technology, and obtaining the internal reference and distortion coefficient of each camera; performing real-time distortion correction on the collected video frames; on the basis of side-by-side splicing or homography matrix perspective transformation, all paths of corrected images are fused into a complete two-dimensional panoramic image, and a view blind area of a single camera is eliminated; calling a lightweight target detection model on the panorama, and extracting bounding boxes and confidence coefficients of all cattle at one time; sorting according to the confidence from high to low, performing IOU non-maximum suppression de-duplication on the detection frame, and ensuring that the same cattle only counts once in an overlapping region; and the whole process is completed through the edge computing node. According to the invention, efficient, real-time and accurate on-site cattle checking of a large-scale cattle farm can be realized.

The invention discloses a deformation monitoring method and system based on a machine vision technology, and relates to the technical field of crossing of structural health monitoring and computer vision, and the method comprises the steps: improving the angular point positioning precision in a calibration stage through a sub-pixel-level Harris angular point detection algorithm; robust recognition and initial positioning of an artificial marker in a first frame of image are realized by using an X-Feat convolutional neural network model, an anti-interference template is generated in combination with histogram equalization and Gaussian filtering, the stability of the system is enhanced, a local search area is delimited by taking an initial positioning point as a center, and a deep learning and interpolation algorithm is combined, so that the anti-interference performance of the system is improved. High-precision tracking and displacement calculation of the marker center are realized, the monitoring resolution is effectively improved, pixel-level displacement is accurately converted into physical space displacement through inverse transformation of a homography matrix, and dual threshold criteria of mode length and change rate are introduced for dynamic early warning, so that the engineering practicability and safety response capability of a monitoring result are improved.

The invention provides a guide registration method, system and device for RGB and infrared image fusion and a storage medium, and belongs to the technical field of image enhancement based on computer vision. Firstly, input of a visible light image and an infrared image is obtained, main target objects in the images are recognized and detected, and a preliminary target detection frame is generated; then, taking the target detection frame as spatial prior information, and generating an RGB image segmentation result and an infrared image segmentation result; performing feature matching on two segmented image results through a newly-designed cross-modal feature matching engine strategy, firstly extracting coarse-grained and fine-grained features through two parallel encoders, then obtaining a homography matrix through a global matcher, distortion refining and matching sampling, and finally obtaining a cross-modal feature matching algorithm; and then homography transformation is carried out on the original infrared image to obtain a final registration image. The method is superior to an existing method in the aspects of registration accuracy and stability, and a robust solution is provided for registration of infrared and visible light images.

This application discloses a fast and interpretable homography decomposition method, relating to the field of computer vision technology. The method includes: obtaining a first anchor point pair, a second anchor point pair, and two sets of auxiliary point pairs based on a source image and a target image; calculating a first similarity transformation matrix based on the first anchor point pair and a second similarity transformation matrix based on the second anchor point pair; the first similarity transformation maps the first anchor point pair to a standard anchor point, and the second similarity transformation maps the second anchor point pair to the standard anchor point; calculating the parameters of a kernel transformation matrix using the auxiliary point pairs on the source plane and the target plane; the kernel transformation matrix has the property of preserving the standard anchor points; and synthesizing a homography matrix from the source plane to the target plane based on the first similarity transformation matrix, the second similarity transformation matrix, and the kernel transformation matrix. This application aims to solve the problems of low computational efficiency, scene limitations, and lack of algebraic and geometric meaning in existing homography calculation methods.

The present application belongs to the field of computer vision and remote sensing photogrammetry, aiming at the problem of being limited by the attitude estimation error of unmanned aerial vehicle and the altitude prior data in the prior art, a target geolocation system and method based on unmanned aerial vehicle image are proposed, the system at least includes a data acquisition module, an end-to-end geolocation module and a benchmark test evaluation module. The method uses the data acquisition module to obtain a data set containing a reference map, a query image and target label information as the data basis and evaluation benchmark of the target geolocation system; the end-to-end geolocation module receives the reference map and the query image, designs a target positioning algorithm, and calculates the geographic coordinates of the target points in the query image; the benchmark test evaluation module quantitatively evaluates the geolocation performance based on the RANSAC homography and the PnP-based pose estimation strategy. The present application improves the geolocation accuracy and reliability by using the system and method.

A computer-implemented method and system relate to an image encoder that receives a digital image as input. The image encoder generates a weight map using the prior feature map. A prior feature map is generated using pixels of the digital image. A weight map is generated based on Lie data associated with the digital image. The homography is interpolated between the two planar projections of the digital image using at least the weight map and the homography matrix. The homography matrix provides a mapping between two planar projections of the digital image. A homography kernel is generated by applying homography to the convolution kernel. A homographic transform kernel is applied to the prior feature map to convolve different planar regions appearing in the digital image, and a new feature map is generated for computer vision tasks involving three-dimensional (3D) perception.

The invention discloses an image matching method, system and device based on open world target detection enhancement, and relates to the technical field of computer vision and image matching. The method comprises the following steps: receiving a to-be-matched first image and a to-be-matched second image, and performing parallel processing; performing dynamic region filtering on bounding boxes and category labels of objects in the first image and the second image; performing category filtering on the initial feature point matching pair between the first image and the second image to obtain a matching set after category filtering; performing bidirectional bounding box projection verification on each matching pair in the matching set after category filtering by using a homography matrix; and performing adaptive matching selection on the matching set after iterative optimization and outputting a final matching set. According to the method, open world target detection is introduced, object category and bounding box information is obtained, and category consistency check and spatial constraint verification are performed based on the object category and bounding box information, so that mismatching of semantic inconsistency and geometric inconsistency is inhibited fundamentally, and the matching precision is remarkably improved.

The invention discloses a method and a system for calibrating a preset position of a pan-tilt camera, belongs to the technical field of computer vision and image processing, and solves the problem that full-process automation and multi-dimensional coordination of image definition optimization, precise pose adjustment and focal length and size calibration cannot be realized in the prior art. According to the method, on the basis of the cradle head camera with the currently set focal length and pose, a gradient-based definition evaluation algorithm is adopted to carry out definition judgment on an original image collected by the cradle head camera so as to obtain a clear image; s2, carrying out feature point matching on the clear image and the standard image, if matching fails, controlling the pan-tilt camera to rotate, executing the step S1 again, and if matching succeeds, carrying out homography matrix solving based on the feature points obtained through matching so as to guide the pan-tilt camera to carry out pose calibration; and after the pose calibration, obtaining a new image to carry out zoom coefficient calculation, and carrying out focal length adjustment according to the zoom coefficient so as to complete final calibration. The method is used for calibrating the preset position of the pan-tilt camera.

The invention relates to the technical field of computer vision nondestructive testing, and discloses a concrete crack nondestructive testing quantitative analysis method based on a convolutional neural network. The method comprises the following steps: firstly, carrying out definition screening on a video stream, and generating a semantic mutual exclusion mask covering a crack region by utilizing semantic segmentation and morphological expansion; removing non-coplanar interference feature points in the crack region under the constraint of a mask, resolving a homography matrix only based on the feature points of the background region, and generating an orthoimage without perspective distortion; and finally, calculating the sub-pixel physical width of the crack in combination with the camera physical distance obtained by homography matrix decomposition and a direction adaptive Zernike moment positioning algorithm. According to the method, interference of crack textures on plane parameter calculation is eliminated through semantic constraint, geometric distortion caused by oblique shooting is corrected through virtual orthographic projection, and high-precision crack width quantitative detection can be achieved without external distance measuring equipment.

The invention discloses a real-time shelf stockout detection method and device based on an inspection robot, and a storage medium, and relates to the technical field of computer vision, and the method comprises the steps: correcting a historical shelf image based on a shelf linear element of a shelf, obtaining a target neural network model through a training data set training model determined based on each corrected image, collecting a current shelf image of a preset point by using a camera on the inspection robot; processing the current shelf image by using the target neural network model to obtain a distorted flow field prediction result, generating a corrected image, feature points and descriptors based on distortion estimation parameters of the distorted flow field prediction result, and matching feature points of an image adjacent to the corrected image based on the feature points and the descriptors to obtain target matching point pairs; and determining a transformed image by using the homography matrix determined based on the target matching point pair and the corrected image so as to determine a vertical frame view, and detecting the vertical frame view to obtain an out-of-stock detection result. Therefore, the efficiency of detecting the goods in the goods shelf can be improved.

This invention discloses a temporal confidence-based dynamic target image recognition method mounted on a sight, specifically relating to the field of computer vision technology. It achieves label-free automatic calibration of the reference benchmark by constructing a normalized virtual pixel coordinate system; it employs a spatiotemporal dual-path parallel architecture to extract target points and calculate semantic matching scores; it establishes a target state monitoring sequence and constructs a temporal confidence evolution model; when encountering instantaneous physical occlusion, it performs nonlinear confidence decay and combines state prediction to maintain tracking, effectively avoiding tracking link interruptions; it extracts the elliptical constraint parameters of the reference image in real time, dynamically updates the homography mapping matrix, compensates for perspective projection distortion of the imaging surface caused by high-dynamic transient disturbances, and achieves sub-pixel-level spatial accuracy correction mapping. This invention solves the technical problems of target recognition interruption and large spatial mapping deviation caused by jitter distortion and short-term occlusion in high-dynamic image acquisition environments, improving the continuity of target recognition.

A computer-implemented method and system refer to an image encoder that receives a digital image as input. The image encoder generates a weighting map using a preceding feature map. The preceding feature map is generated using pixels of the digital image. The weighting map is generated based on Lie data associated with the digital image. A homographic transformation is interpolated between two planar projections of the digital image using at least the weighting map and a homography matrix. The homography matrix provides a mapping between the two planar projections of the digital image. Homographically transformed kernels are generated by applying the homographic transformation to convolution kernels.The homographically transformed kernels are applied to the preceding feature map to perform folding on different planar areas appearing in the digital image and to generate a new feature map that is used for a computer vision task involving three-dimensional (3D) perception.

The application discloses a kind of mobile robot storage location identification method based on computer vision, it is related to machine vision technical field, including, acquisition continuous image and distortion removal, obtain luminance diagram and effective pixel mask;In frame window, the luminance sequence is counted according to line and low frequency is removed, and the frequency of LED stroboscopic frequency, modulation amplitude, intra-frame time offset and row-by-row time increment are estimated based on the joint estimation of row sequence period term and cross-frame phase difference, and the row-level modulation term is calculated;Accordingly, the luminance diagram of each frame is normalized according to line, key points and straight line segment candidates are extracted, and spatial consistency tracking is carried out through interframe homography mapping, and a stroboscopic robust key point set and a stroboscopic robust straight line set are obtained;Further estimate the main direction of the shelf and plane shaping, and the grid lines form the grid candidate area, determine the storage location number in combination with the sign analysis, and obtain the storage location pose by solving the pose from the grid corner and the physical size of the grid.

The invention belongs to the technical field of computer vision, and provides a lane line constraint-based BEV conversion method without internal and external parameters. The method comprises the following steps: proposing a homography matrix calculation model of lane line constraints under the condition that proper internal reference setting is determined, and calculating a homography matrix analytical solution by utilizing one or two groups of lane line parallel constraints; an optimization solution method based on multi-group lane line input is provided, and the robustness of conversion in an automatic driving scene is improved; a processing assembly line is provided to input a picture, detect a lane line and output a BEV image. According to the method, the problem that BEV view angle picture generation depends on internal and external parameters can be solved, and a reliable and robust method is provided for homography matrix calculation of view angle conversion.

The invention discloses a meal delivery detection method and device based on computer vision and a storage medium, and relates to the technical field of computers, and the method comprises the steps: carrying out the first segmentation operation of a target meal delivery reference image through a preset instance segmentation model, carrying out the frame extraction of a meal delivery video stream, and carrying out the second segmentation operation of an obtained real-time image; based on the obtained second segmentation result, judging whether a meal is in a meal delivery state, and if the meal is in the meal delivery state, determining a first dinner plate corner point and a second dinner plate corner point based on a first mask in the first segmentation result and a second mask in the second segmentation result and by utilizing a preset minimum bounding rectangle function; and respectively correcting the target meal delivery reference image and the real-time image by using the homography matrix, and carrying out comparison judgment on the quantity of dishes and comparison judgment on the positions of the dishes by using the corrected reference image and the corrected real-time image so as to obtain a comprehensive detection result of whether the meal delivery of the real-time image meets a preset meal delivery condition or not. The accuracy of meal delivery detection is improved.

The application discloses a computer vision-based dish delivery detection method and device and a storage medium, and relates to the technical field of computers, which comprises the following steps: performing a first segmentation operation on a target dish delivery reference image by using a preset instance segmentation model, performing frame extraction on a dish delivery video stream, and performing a second segmentation operation on the obtained real-time image; judging whether a dish delivery state is present based on the obtained second segmentation result; if the dish delivery state is present, determining first and second dish plate corner points based on a first mask in the first segmentation result and a second mask in the second segmentation result and by using a preset minimum circumscribed rectangle function; correcting the target dish delivery reference image and the real-time image by using a homography matrix, and performing comparison judgment on the number of dishes and the position of dishes by using the obtained corrected reference image and corrected real-time image, so as to obtain a comprehensive detection result of whether the dish delivery of the real-time image meets preset dish delivery conditions. The accuracy of dish delivery detection is improved.

The invention discloses a fire safety monitoring system based on computer vision, particularly relates to the field of computer vision, and is used for solving the problems of low fire source identification precision, large illumination interference and difficulty in unified analysis of multi-view monitoring images in existing fire monitoring. A candidate hot spot extraction module is arranged to perform binarization processing on a monitoring image to extract a candidate hot spot region, a morphological enhancement module is utilized to adaptively adjust a nuclear parameter fusion hot spot region, and a standard view angle correction module is utilized to realize perspective correction based on feature point detection and homography matrix calculation; when a multi-view-angle image exists, a view angle fusion module is used for calculating three-dimensional point cloud data of a hot spot based on epipolar geometry, and a real hot spot is identified through time sequence and space consistency verification; and finally, the monitoring output module completes spatial positioning and risk extension trend prediction of real hot spots, outputs fire risk position and trend information, and realizes high-precision fire source identification and real-time early warning in a complex scene.

The application belongs to the technical field of computer vision and image processing, and discloses an occlusion detection method and system based on image registration and depth map and a storage medium, and the main idea is to solve the false detection problem caused by light change or local texture change. The main scheme comprises the following steps: matching a reference image and a to-be-detected image, calculating a homography matrix to align the images through RANSAC after filtering outliers, generating a three-channel residual image and converting it into a gray image. Morphological opening operation and binarization are performed on the gray residual image, and the connected regions with an area exceeding a threshold value are selected as candidate occlusion regions. Depth maps of two images are generated and aligned by using the homography matrix, the depth ratio is calculated after eroding the non-candidate regions, and the depth scales of the to-be-detected image and the reference image are unified. The average depth difference of the candidate regions in the two images and the average distance from the camera optical center plane are calculated, and if the depth difference exceeds a threshold value and the distance meets the conditions, the real occlusion is determined. The area proportion of the real occlusion region is counted, and if it exceeds a threshold value, the occlusion warning is triggered.

The invention discloses an interest point detection model training method, a product positioning method, equipment and a medium, and relates to the technical field of computer vision, and the method comprises the steps: obtaining a sample image pair; extracting first information of each interest point in the original image through a to-be-trained interest point detection model, and extracting second information of each interest point in the transformed image; performing homography transformation on the position of each interest point in the original image, matching with the position of each interest point in the transformed image to obtain each interest point pair, and matching with the position of each interest point in the original image to obtain each interest point pair; calculating the sample loss of the sample image pair based on the first information and the second information corresponding to each interest point pair; and adjusting parameters of the interest point detection model based on the sample loss of the plurality of sample image pairs. According to the unsupervised interest point detection model training method provided by the invention, the method does not need to depend on any manual annotation data, and the training efficiency and the adaptive capacity of the model in different scenes are improved.

This invention belongs to the field of computer vision technology and proposes a lane-constrained BEV conversion method that does not require internal and external parameters. This method proposes a lane-constrained homography matrix calculation model under appropriate internal parameter settings, and uses one or two sets of lane-line parallel constraints to calculate the analytical solution of the homography matrix. An optimized solution method based on multiple lane line inputs is proposed to improve the robustness of the conversion in autonomous driving scenarios. Finally, a processing pipeline is proposed to detect lane lines from an input image and output a BEV image. This invention can solve the problem of BEV perspective image generation relying on internal and external parameters, providing a reliable and robust method for calculating the homography matrix for perspective conversion.

The invention discloses a zoom camera self-calibration method and device based on gravity constraint, and relates to the field of computer vision. The method comprises the following steps: acquiring two images with different focal lengths and corresponding gravity vectors; aligning a camera coordinate system by using a gravity direction; matching at least three pairs of feature points, and correspondingly parameterizing the two-dimensional homography matrix; establishing a mathematical relationship among the two-dimensional homography matrix, the two unknown focal lengths and the Euclidean homography matrix; constructing a polynomial equation set by using constraints derived from the hypothesis that the ground plane is orthogonal to the gravity direction; and finally solving the polynomial equation set to obtain two focal length values. According to the method, the variable focal length of the camera can be solved only by corresponding at least three pairs of feature points, the calculation efficiency is high, the method is suitable for applications with high real-time requirements, meanwhile, physical constraints (the gravity direction) and scene priori (the ground plane is orthogonal to the gravity direction) are introduced, and the robustness of the algorithm and the precision of the solution are guaranteed.

The present application relates to the technical field of computer graphics and computer vision, and discloses a homography matrix estimation method and system, the method comprising the following steps: S1, collecting images with different overlap rates as a training data set; S2, constructing a neural network model, including an overlap monitoring network and a homography matrix estimation network; S3, using a loss function to guide and optimize the neural network model, calculating the loss value between the result of image transformation by the network output homography matrix and the target image; S4, training the neural network model to generate a trained neural network model; S5, using the overlap detection network to crop out the common area of the two estimated images; S6, using the deep homography matrix estimation network to achieve coarse-to-fine homography matrix estimation. The system comprises an acquisition unit, a model construction unit, a model training module and a homography matrix estimation module. The present application also discloses an electronic device and a computer readable storage medium.

This invention discloses a depth homography estimation method based on dense scene matching, belonging to the field of computer vision. The method proposes a coarse-to-fine depth homography estimation network consisting of three sub-models, combining linear and nonlinear homography estimation. The first two sub-models perform linear homography estimation for coarse alignment of two images, while the last sub-model performs nonlinear homography estimation for fine alignment through mesh deformation. This coarse-to-fine estimation approach further improves image alignment accuracy within a limited network capacity. Furthermore, the method explicitly estimates corner offsets through a dense scene matching module, considering the correlation between image feature points and incorporating matching information. This more reasonably establishes the connection between the image and corner offsets, avoiding network overfitting and improving the network's ability to process image details.

The invention discloses a dual-light image nonlinear registration method and system for multi-source fusion monocular depth estimation, and relates to the technical field of computer vision and image processing. The method comprises the following steps: constructing a mapping relation library of a distance and a homography transformation matrix; an improved target detection and depth perception integrated model is adopted, detection information, depth probability distribution and confidence of a to-be-detected target are synchronously output, geometric depth estimation based on the real physical size of the target is combined, and self-adaptive weighted fusion is carried out by utilizing the confidence to obtain a final target distance; and then table look-up and interpolation are carried out in a mapping relation library based on the distance to generate an adaptive homography transformation matrix, and target coordinates in the first spectral image are accurately mapped to the second spectral image. According to the invention, through fusion of geometric prior and deep learning, adaptive nonlinear registration of targets at different distances is realized, and the registration precision and robustness of the non-coaxial dual-light system are significantly improved.

The invention relates to an unmanned aerial vehicle yaw angle estimation method based on image feature matching, and belongs to the technical field of image processing and computer vision, and the method comprises the steps: inputting a satellite remote sensing image and a squint image shot by an unmanned aerial vehicle for preprocessing, and calculating the average texture intensity of the image; according to the average texture intensity, selecting an applicable feature extraction algorithm to perform feature point extraction, and obtaining feature point descriptors of the two images; matching and screening feature point descriptors in the feature point descriptor set to obtain a plurality of groups of reliable matching point pairs, and calculating a homography matrix; according to re-projection errors of the matching points and the homography matrix, error points in the matching point pairs are eliminated, and an accurate homography matrix is obtained; related information is extracted from the homography matrix, the horizontal angle of the unmanned aerial vehicle relative to the target area is calculated, and the yaw angle of the unmanned aerial vehicle is calculated in combination with the rotation angle of the camera holder. Auxiliary support is provided for positioning, navigation and rapid deployment of the unmanned aerial vehicle.

This invention relates to a method for tracking basketball players and obtaining their trajectory coordinates based on a visual fundamental model, belonging to the field of computer vision and intelligent sports analysis technology. The method includes: S1, acquiring a video stream and calculating a homography matrix using key semantic points; S2, performing global color clustering on the image to generate adaptive text prompts; S3, constructing a hybrid inference pipeline, using the prompts to drive the visual fundamental model to perform full-image concept segmentation on keyframes to obtain pixel-level masks, with interpolation used for intermediate frames; S4, traversing the mask and locating the geometric center of the bottom pixel row as the precise contact point; S5, mapping the contact point to physical coordinates using the homography matrix and associating it with identity based on color and mask features; S6, constructing a conditional diffusion probability model to perform reverse denoising and reconstruction on noisy trajectories, outputting smooth trajectory data. This invention achieves zero-shot generalization and high-precision trajectory acquisition through a visual fundamental model and mask contact point extraction.

The invention relates to the technical field of computer vision and positioning, in particular to a tight coupling method for vision relative positioning and absolute positioning. The technical problems that in the application process of a visual positioning method in the prior art, the visual positioning method depends on a holder, information utilization is insufficient, and the adaptive capacity is poor are solved. According to the technical scheme, the tight coupling method for visual relative positioning and absolute positioning comprises a visual relative positioning step, a reference image obtaining step, a real-time image preprocessing step, a feature matching step and an optimization step, in the feature matching step, feature matching is conducted on a preprocessed real-time image and a reference image, a homography matrix is calculated, and the real-time image and the reference image are optimized; mapping feature points on the real-time image to a world coordinate system in the visual relative positioning step; according to the method, visual absolute positioning and relative positioning are tightly coupled under a factor graph framework, and meanwhile, position, attitude and scale parameters are optimized, so that high-precision and high-robustness positioning is realized.