56results about How to "The result is robust" patented technology

The invention discloses a clothes making method based on three-dimensional to two-dimensional clothes pattern seamless mapping. The clothes making method based on the three-dimensional to two-dimensional clothes pattern seamless mapping comprises the following steps: constructing a three-dimensional clothes model, and establishing a mapping relationship between the three-dimensional clothes model and a two-dimensional sample plate mesh model; designing a clothes pattern on the surface of the three-dimensional clothes model; calculating a pattern on the seam allowance of a two-dimensional sample plate; and carrying out digital printing, cutting and sewing of the two-dimensional sample plate to make real clothes. According to the clothes making method based on the three-dimensional to two-dimensional clothes pattern seamless mapping, the problems that registration and grid matching are complex, and seamless splitting can not be realized since the pattern continuity of the clothes is damaged in a conventional clothes making process at present are solved, the pattern of the clothes is more beautiful and intact when people wear the clothes, cloth is saved, and time and cost for registration and grid matching in the clothes making process are respectively shortened and reduced. The clothes making method based on the three-dimensional to two-dimensional clothes pattern seamless mapping has the advantages of specific algorithm, friendly interface and robust result.

A super-pixel-level image global matching method comprises the steps of obtaining an input image pair corrected by polar lines through a binocular stereo camera, calculating a self-adaption cross of each pixel of the input image pair so as to obtain a self-adaption window of the current pixel, calculating the matching cost of the pixel, adopting a replacement strategy to process a coverage area and adopting a suboptimum strategy to process an image boundary; establishing super-pixels for images, conducting planar fitting on a parallax value of each super-pixel area to determine reliable pixels and deleting obvious wrong planes so as to determine initial parallax plane set; calculating the matching cost of the super-pixels according to the obtained matching cost of the pixel, establishing a data item and a smoothing item and utilizing a Graph-Cut optimization algorithm to conduct continuous iteration on an energy equation so as to obtain a final parallax plane. The super-pixel-level image global matching method can effectively avoid the problem that image noise, distortion or pixel value abnormity and other situations easily occur in a weak texture area, a discontinuous-parallax area and the coverage area, is good in robustness and can obtain depth information more approximating to real scenes.

The invention discloses a feature-based real-time video image abstraction method, which comprises five steps of converting RGB color space of the real-time videos and images into CIE-Lab color space, adopting a bilateral grid to carry out the rapid bilateral filtering of a brightness channel, carrying out the pseudo-quantization of the brightness value, adopting the feature-flow-based anisotropic Gaussian differential line to extract the algorithm extraction line region and carry out the combined operation of the extracted line region and the similar brightness value to strengthen the lines of the line region and re-converting the real-time videos and images to the RGB color space from the CIE-Lab color space. Each step of the invention can be parallelly processed in a graphics card hardware of a household computer, thus effectively solving the rigid and incoherent problems of the extracted lines in the existing abstraction method of multiple real-time videos and images, and providing the simple, convenient, intuitive and real-time video and image abstraction method for non-professionals.

The invention discloses a cardiac ultrasound imaging method, and aims at solving the problem of automatically segmenting of a left ventricle contour during a cardiac measurement process, and preventing the situation of inaccurate volume calculation which might be caused by manually specifying left ventricular segmentation points in the measurement process. Compared with a traditional left ventricular segmentation algorithm, a large number of images of the position of a left ventricle manually marked by a doctor are used to train a deep convolutional neural network, the position of the left ventricle can be accurately identified by the acquired detection network, and a level set algorithm can be performed within a detection position area to obtain a segmentation curve of a left ventricularendocardium; compared with a traditional algorithm, acquired results are more robust, the operation efficiency is higher. According to the cardiac ultrasound imaging method, a manual enhancement manner is adopted to enhance edge areas to improve a segmentation effect aiming at the situation of poor image quality of ultrasound mapping.

The invention discloses an automatic adjustment method of a three-dimensional human face model, comprising the following steps: a three-dimensional scanner obtains a three-dimensional human face modeland obtains a projection map of the three-dimensional human face with texture from three angles; the three-dimensional scanner obtains a projection map of the three-dimensional human face with texture from three angles; 2D feature point detection of projection map; 2D feature point are backprojected into 3D to obtain 3D feature point; The PCA facial model is used to fit the original 3D facial model. 3D registration of that initially align mesh is performed to solve the energy equation; Recalculating texture coordinates for the final aligned mesh; After the registration of the three-dimensional face adjustment, get the adjusted three-dimensional face model. The invention adopts a set of technology for automatically recognizing three-dimensional human face characteristic points and adjusting the human face on the basis of the technology, It can beautify 3D face captured by 3D scanner with one button, simple and convenient, and has clear algorithm, friendly interface and robust results.It can be used in virtual face plastic surgery, computer animation and computer games and other fields.

The invention discloses a reconstruction method based on point-line feature rapid fusion. The method comprises the following steps of intercepting a video as an image, and performing preprocessing, such as focal length extraction, downsampling, etc., to reduce the reconstruction complexity; carrying out the point feature matching, carrying out the point feature extraction and matching by adoptingscale invariant feature transformation; matching the line features quickly from coarse to fine, using the line segment segmentation detector features for extracting and describing the line features, and obtaining an image line segment feature matching pair through the four steps of violent matching, motion estimation, Hamming distance threshold judgment and length screening on the line segment descriptors; carrying out the point-line feature fusion, converting a final line segment feature matching pair into the pixel points, analyzing the pixel points and the pixel coordinate positions of theexisting point features, deleting the repeated points, and then fusing the line segment pixel points and the point features; calculating a camera outer posture and a three-dimensional point cloud, calculating a substantive matrix by using a final image point-line feature matching pair, solving the camera outer posture, solving the three-dimensional point cloud in a triangularization manner, and optimizing a result by using a beam adjustment method.

The invention discloses a diffusion curve-based RGBD image vectorization method. The method includes the following nine steps: inputting an original RGB color image and a depth image D which are to be processed, performing multi-scale Canny edge extraction on the RGB image, coloring the acquired multi-scale binary edge image to generate a color edge image, restoring the depth image, performing depth edge extraction on the restored depth image D' to generate a depth edge image, performing subtraction on the two edge images to obtain a detail edge image, performing tracking merging on the detail edge image and the depth edge image to generate a group of broken line sections, performing color sampling and Bezier curve fitting on the broken line sections to obtain a group of diffusion curves, using colors on the curve as constraints to solve a Poisson's equation to obtain a vectorization result. The diffusion curve-based RGBD image vectorization method provided by the invention adopts RGBD images to obtain an object outline, well restores the real outline of an object, and solves the situation of multi-scale Canny invalidation in some color environments. The method has a clear algorithm and a robust result, and is suitable for vectorization of the RGBD images.

The invention discloses a method for abstracting real-time videos and images based on visual attention and comprises the following seven steps: inputting the real-time videos and images to be processed, converting red, green and blue color spaces of the real-time videos and images into a CIE-Lab color space, generating a visual attention scatter diagram, deriving a functional diagram of an interest region, adopting an abstracting method based on a characteristic flow field to obtain initial real-time videos and images, adopting linear interpolation technology to obtain new real-time abstracting videos and images, and reconverting the real-time video and the images from the CIE-Lab color space into the red, green and blue color spaces. Each step of the method can be processed concurrently in graphics card hardware of a home computer; the method effectively solves the problem that the prior real-time video abstracting method has even abstract; and the invention provides a simple and intuitive abstracting method for real-time videos and images for laypeople.

The present invention discloses a 3D portrait food printing method. The method comprises the following steps: a human face image is input, the input human face image is convert to a gray scale map, 83 human face feature points are found, preprocessing is conducted, gray values of the human face feature points are strengthened, an ETF algorithm is used to construct a smoothing direction field of the human face image, a FDOG algorithm is used to conduct anisotropic Gaussian filter based on a direction field flow direction, binarization is used to obtain human face image contours, an 8-connected depth-first traversa is used to get rid of local details to obtain simplified human face contour image, a prim algorithm is used to connect two closest line segments to form an one-stroke human face contour path, an erosion algorithm is used to simplify the path to finally obtain a human face 3D printing path. The 3D portrait food printing method simplifies the portrait 3D printing generation path, at the same time retains most contour information on human face, is also clear, friendly in interface is, and great in results, and can be used for the 3D food printing, commercial apps and other fields.

The invention relates to the technical field of medical ultrasonic imaging, in particular to a quantitative shear wave elastography method. According to the ultrasonic quantitative elastography method and system based on strain of sliding window linear fitting and a shear wave velocity detection algorithm using two-dimensional linear fitting, higher noise immunity can be achieved, and a more reliable result can be obtained; meanwhile, under the condition that the load of an ultrasonic front-end storage and transmission module is not increased, ultrasonic whole-domain quantitative elastography is achieved, and the design difficulty of the ultrasonic quantitative elastography system and the equipment cost are greatly lowered.

The invention discloses a method for point-to-point matching between non-rigid three-dimensional models. The method comprises the following steps: firstly, establishing an anisotropy spectrum manifoldwavelet descriptor; and taking the established descriptors as descriptor constraints of model points, and taking a thermonuclear relationship of each point on the model as a point pair relationship constraint to establish an objective function so as to realize optimal matching between the model points. According to the method, an anisotropy spectrum manifold wavelet descriptor is established at the earlier stage, and then a thermonuclear relationship is adopted as a point pair relationship constraint. Compared with an existing method, the anisotropy spectrum manifold wavelet descriptor has the advantages of being invariant in isometric deformation, capable of distinguishing intrinsic symmetry of the model, high in resolution capability and positioning capability, high in calculation efficiency and compact in structure. The thermonuclear relationship serving as a point pair relationship constraint is more excellent in calculation efficiency and stability than other methods adopting geodesic distance. Therefore, the method is clear in calculation, robust in result and accurate in matching.

The invention discloses a marble vein glaze pattern real-time generating method, which comprises the following steps of: appointing the size of a vein pattern and the color of a background pattern layer; building a base pattern by means of dropping, spraying, overturning and the like; building a marble vein glaze pattern by means of user operations such as comb, stylus, wave, rotation and the like; and multiplying the density field of each layer by the color of the corresponding pattern layer and adding to obtain a final display color. The method overcomes the defects which are commonly existed in a computer simulation system of the existing marble vein glaze pattern technology such as slower speed, fuzzy generating pattern, unfriendly user interface, over-abstract physical equation parameter and the like, adopts the real-time marble vein glaze pattern generating method based on mathematical function, and guarantees that the generated marble vein glaze pattern has the characteristics of the clear boundary and the elegant pattern of the real marble vein glaze pattern by adopting the density field expressing method by means of reverse mapping, supersampling and multi-pattern layer and the iterative density field updating arithmetic.

The invention discloses an ellipse detection method based on chord tangent distance. The method comprises the following steps: 1, image edge pixel chain extraction: converting an input image into an LAB color space, extracting an image edge on an L color channel, performing connected component marking and edge tracking operations, and obtaining an edge pixel chain from each connected component; 2,ellipse fitting: under a least square fitting framework, establishing an objective function by using a chord tangent distance as a fitting error metric, and iteratively solving ellipse parameters byusing a Gauss-Newton method. Due to the fact that the chord tangent distance is adopted, a simple closed-form solution is provided for calculating the corresponding point of one point on the given ellipse. The iterative solution does not need to be carried out like most of existing geometric distance fitting methods. The calculated amount is greatly reduced. Meanwhile, errors occurring in calculation are reduced, and the practicability of the method is improved.

The invention relates to big data processing, and aims to provide a power sales amount intelligent prediction method based on a deep recurrent neural network. The method comprises the steps of readinghistorical data of sales flow and electricity consumption of an electric power department, performing information mining and analysis after preprocessing, and evaluating a relation between the amountpayment time and user payment time to obtain distribution information; organizing a historical data structure, taking normalized n-day data as input, learning high-dimensional characteristics by using a multi-layer recurrent neural network (GRU), and inputting the high-dimensional characteristics into a softmax discriminator to carry out sale amount level classification in a certain period of time; and traversing hyper-parameters of the deep circulation network model by using a grid method, recording the optimal hyper-parameters after multiple experiments, constructing a final amount prediction deep circulation neural network model, and intelligently predicting the electric power sales amount by using the final amount prediction deep circulation neural network model. The method is more accurate and reasonable, manual intervention is less, the result is more robust, the method is more suitable for big data, and automatic learning can be realized.

The invention discloses a method for accelerating decoding of a camera video stream through a programmable GPU in a virtual-real fusion system. The method comprises the steps of obtaining an originalYUV video; Creating a single-channel texture for the pixel flow of each frame of YUV image and transmitting the single-channel texture to the GPU; Drawing a three-dimensional virtual scene, and judging whether the current pixel belongs to a camera image or not; And mixing the RGB pixels and the three-dimensional scene to form a final virtual-real fusion image. Compared with a traditional method for decoding YUV video pixels on a CPU in a multi-thread mode, more original video pixels which do not need to be sampled can be effectively cut, and meanwhile the decoding efficiency is greatly improved on the basis of a GPU with higher parallel decoding capacity. The method is clear in algorithm and robust in result, can be applied to fusion application of a large three-dimensional virtual scene and real-time monitoring, and can be well combined with a video of a current mainstream camera manufacturer for use.

The invention discloses a sample evaluation method of a coronary artery segmentation model, comprising the following steps: S1, inputting a sample to train a 3D coronary artery segmentation model to obtain volume prediction data; S2, slicing the sample and the corresponding volume prediction data along the Z direction, and calculating the score of each slice; S21, marking the sample slice; S22, comparing the sample slice mark result with the corresponding volume prediction data slice result to obtain the score of the sample slice; S23, taking the mean value of the scores of each sample slice to obtain the scores of the samples. At the same time, the invention also discloses two training methods of the coronary artery segmentation model, namely, utilizing the sample score to evaluate the samples of the coronary artery segmentation model for single training or multiple training respectively, and endowing different training strategies according to the difficulty degree of the samples, soas to obtain a better training model.

The embodiment of the invention relates to the field of image processing, and discloses a bone segmentation method in a hip joint image, electronic equipment and a storage medium. The method comprises the steps: obtaining a to-be-segmented hip joint image; inputting the to-be-segmented hip joint image into a pre-trained segmentation model, and outputting a segmentation result, wherein a method for obtaining the segmentation model through pre-training comprises the following steps: creating an initial segmentation model, and obtaining a plurality of artificially labeled hip joint sample images to obtain a mask image; inputting the hip joint sample images into a self-attention transformation initial model and a convolutional neural network initial model to respectively obtain a first segmentation result and a second segmentation result; and calculating training loss, and returning the training loss to the initial segmentation model to obtain a final segmentation model. According to the embodiment of the invention, the segmentation result is accurate, the robustness is realized, and bone structures in the hip joint images can be efficiently and automatically segmented, so that clinical doctors are assisted in surgical planning, intraoperative navigation and postoperative evaluation.

The invention provides a modeling method based on convolutional curved surface approximation in the VR environment. The method comprises steps that one type of point, line and plane frameworks is selected, sizes and thickness are set, the data is inputted to the framework through a button of a VR handle, a convolutional curved surface is utilized to carry out approximation of the point, line and plane polygon frameworks having the radius information, the acquired input framework is transferred to a GPU through a system to carry out parallel field calculation and extract contour surfaces. The method is advantaged in that problems of complex operation caused by modeling generally based on two-dimensional planes in a 3D modeling tool in the prior art and mismatching between modeling input sizes and output model sizes in the VR are solved, a convolutional curved surface approximation method based on the frameworks is utilized, for approximation of the different-diameter frameworks inputted by users in an interaction system, closed solution schemes are derived respectively for line segment framework tail ends and polygon framework edges. Moreover, for various complex shapes, different humanized coating schemes are provided for simplified input.

The invention provides a priority and multi-granularity based content sharing method under an opportunity network environment, and relates to the content sharing method of an opportunity network. The content sharing method comprises the following steps: 1), defining shared content information and request information including granularity by a user, and calculating a node friendship degree; 2), establishing a communication link and exchanging the respective request collection and the carried content segment statement; 3), calculating the priority of a node storage content; 4), if the communication link still exists, then copying the content segment with the highest priority to a forwarding queue from a cache by the node, or ending the current encountering communication opportunity; 5), if the receiving node has enough storage space, then transmitting the content segment to a counter node by the node to realize the distribution and propagation of the shared content; otherwise performing cache management on the receiving node, that is converting the content segment of high granularity according to the importance weight, so as to make room for storage to perform data transmission; 6), if the node still has a content block which is set with priority, then turning to step 4); otherwise disconnecting the current link, and waiting for the next opportunity of communication and content transmission.

The invention discloses a projection mapping method of a video in a three-dimensional scene based on a screen space, comprising the steps of preprocessing to obtain internal and external parameters ofa projection source and a depth map; Drawing three-dimensional scene from current viewpoint, obtaining diffuse reflection map and depth map of current viewpoint; The texture is drawn from the projection source, and each pixel of the projection picture is converted to the screen space of the viewpoint for projection calculation. The problem of occlusion penetration caused by the loss of depth value in traditional projection texture mapping is solved, the method not only enhances the rendering effect significantly, but also reduces the number of rendering batches and enhances the rendering efficiency. The invention has the advantages of clear algorithm, robust result and high operation efficiency, and is very suitable for application in real-time rendering system. The method can be well combined with a large-scale three-dimensional video monitoring system.

The invention relates to Kongming lanterns, in particular to a computer aided Kongming lantern manufacturing method. The method includes: selecting a required Kongming lantern three-dimensional model from a model library; cutting open the bottom of the Kongming lantern three-dimensional model to sever as an opening of the Kongming lantern; allowing the bottom opening to be close to a circle so as to conveniently attach the bottom opening to a bamboo frame, removing high-frequency noise of the Kongming lantern three-dimensional model, and improving volume-area ratio of the three-dimensional model to allow the three-dimensional model to rise easily; using unfoldable curves and planes to approach the original three-dimensional model, and keeping feature correspondence to the original three-dimensional model during approaching as much as possible; determining minimal size, capable of rising successfully, of the shape-fixed Kongming lantern according to related parameters; unfolding all unfoldable pieces of the Kongming lantern to a two-dimensional plane, printing and outputting to flame-retardant paper, cutting according to the printed shape, assembling the cut pieces into an air sac of the Kongming lantern, sleeving the bamboo frame, and placing fuel. By the method, armatures can manufacture exquisite Kongming lanterns of different shapes and capable of rising.

The invention relates to a semi-online machine-set multi-target tracking method, which comprises the following steps of: obtaining a detection frame of a pedestrian or a moving target according to a pedestrian or moving target video, obtaining a Kalman sequence spectrum according to position change information between the detection frames in a period of time window, finding a pair of Kalman headsaccording to the Kalman sequence spectrum, and tracking the pedestrian or the moving target according to the pair of Kalman heads. obtaining a detection frame of a target or a moving object to be tracked in the next frame through the similarity of the appearance model, the moving model and the size change model, enabling the target or the moving object to be located in the detection frame in the frame, and otherwise, indicating that the target is lost; and splicing the detection frames of which the similarity is higher than a threshold value into the Kalman sequence spectrum, updating the motion model and the appearance model in the Kalman sequence spectrum, and tracking the pedestrian or moving object target in the next frame. The method is suitable for any trajectory splicing type multi-target tracking algorithm, that is, constraint of different trajectories generated by multiple targets such as pedestrians and moving objects is avoided, the tracking precision can be effectively improved, and the identity conversion value is reduced.

The invention discloses a feature-based real-time video image abstraction method, which comprises five steps of converting RGB color space of the real-time videos and images into CIE-Lab color space, adopting a bilateral grid to carry out the rapid bilateral filtering of a brightness channel, carrying out the pseudo-quantization of the brightness value, adopting the feature-flow-based anisotropicGaussian differential line to extract the algorithm extraction line region and carry out the combined operation of the extracted line region and the similar brightness value to strengthen the lines of the line region and re-converting the real-time videos and images to the RGB color space from the CIE-Lab color space. Each step of the invention can be parallelly processed in a graphics card hardware of a household computer, thus effectively solving the rigid and incoherent problems of the extracted lines in the existing abstraction method of multiple real-time videos and images, and providingthe simple, convenient, intuitive and real-time video and image abstraction method for non-professionals.

The invention discloses a Hyaluronic acid injection beautifying method in a three dimensional face virtual plastic surgery; the method comprises the following steps: 1, using a three dimensional scanner to obtain a three dimensional face model, and enabling a user to select injection parameters that includes an injection point, an injection radius and an injection capacity; 2, using a breadth-first traversal method to find all points in an injection area; 3, using a distance field function to calculate the weight of each point in the injection area; 4, using a dichotomy iterative method to calculate the elevation distance of a summit according to the injection capacity inputted by the user, thus obtaining a deformed three dimensional face model. The method uses the three dimensional modeldeformation technology, can simulate the Hyaluronic acid injection effect on the three dimensional face model gathered by the three dimensional scanner, is simple and easy to use, clear in algorithm,friendly in interface, robust in result, and can be applied in virtual face lifting, computer animation and computer gaming field.

The invention, which relates to the image processing and application field, provides a cartoon image compression method based on explicit hybrid harmonic diffusion. The method comprises the steps of feature line extraction, feature line position coding, image color coding, feature line position decoding, and image color decoding. In order to solve problems of poor practicability and the like due to limited color expression capability and long decoding time of the existing partial-differential-equation-based second-generation image compression algorithm, the invention provides a cartoon image compression method based on explicit hybrid harmonic diffusion. With simultaneous utilization of harmonic and biharmonic diffusion, diversified color changes in the image can be coded well, thereby avoiding solution of a large linear system during decoding and realizing real-time decoding. Explicit approximation is carried out on the biharmonic process by using a green function, thereby improving the decoding speed substantially. Moreover, the algorithm is clear; the effect is obvious; the real-time performance is high; and the result is robust. After industrialization, the business value of the unit flow in the mobile phone animation industry can be improved substantially and the user experience is also improved.

The invention discloses a three-dimensional face model adjusting method, which comprises the following steps: appointing at least 55 initial characteristic points on a three-dimensional face model and demarcating 6 characteristic areas; carrying out symmetrization adjustment locally and globally on the initial characteristic points; then globally adjusting the proportion of the symmetrically adjusted three-dimensional characteristic points; and carrying out nonlinear least squares optimization on the proportionally adjusted three-dimensional characteristic points to adjust the side face and adopting a Laplace morphing algorithm to solve the linear system of equations of the initial characteristic points and the three-dimensional characteristic points after side face adjustment so as to obtain the adjusted three-dimensional face model. The invention adopts a Laplacian morphing technology maintaining detailed characteristics so as to retain most detailed characteristic information on the face at the time of adjustment and beautifying, the algorithm is clear, the interface is friendly, and the result is robust, so the method can be used in the fields of virtual face lifting, network video chat, computer game, video conference, computer animation and the like.

The invention provides an animation image adaptation method with controllable redundancy and relates to animation content terminal adaptation of computer assistance. The method comprises the steps that 1), an input animation image is analyzed, and feature lines are extracted; 2), the redundancy is calculated according to the distribution condition of the feature lines to generate a redundancy image; 3), importance degrees of the animation image are calculated; 4), the boundary of a redundancy region is determined according to the redundancy, and a triangular grid is generated with the image boundary, the feature lines and the boundary of the redundancy region as constraint; 5), according to the screen resolution of a mobile terminal, a triangular grid of a non-redundancy region is properly deformed according to the important degree difference so that the whole grid can adapt to a screen; 6), the redundancy region is updated by adopting a mode of barycentric coordinate interpolation, and when the redundancy is larger than a set threshold value, the feature lines can be combined automatically; 7), the grain mapping function of a graphic card is utilized for embedding the image into the triangular grid, and the image is updated automatically along with changes of the grid. The algorithm is definite, an interface is friendly, and a result is robust.

The invention particularly relates to a single-model fusion method based on cross validation. The single-model fusion method based on cross validation adopts a basic machine learning method and comprises the steps of data preprocessing, data set division, cross validation, model training, prediction and model fusion. According to the cross validation-based single model fusion method, cross validation is utilized to improve model generalization, and a multi-model fusion idea is combined to mine data at different levels. Meanwhile, a model is relatively single. A result is generally relatively robust, and the prediction effect is superior to that of a single model or a common multi-model fusion method.

The invention discloses a method for accelerating the decoding of a camera video stream by a programmable GPU in a virtual-real fusion system, comprising: obtaining original YUV video; creating a single-channel texture from the pixel stream of each frame of YUV picture and transmitting it to the GPU; drawing a three-dimensional virtual scene, and judging Whether the current pixel belongs to the camera image; RGB pixels are mixed with the 3D scene to form the final virtual-real fusion image. Compared with the traditional method of decoding YUV video pixels in multiple threads on the CPU, the present invention can effectively cut more original video pixels that do not need sampling, and at the same time, based on the GPU with stronger parallel decoding capability, the decoding efficiency can be greatly improved. The algorithm of the invention is clear and the result is robust, and the method can be used in the fusion application of large-scale three-dimensional virtual scene and real-time monitoring, and can be used in good combination with the current mainstream camera manufacturer's video.