CN-121999143-A - Incremental SfM reconstruction method based on relative orientation uncertainty

CN121999143ACN 121999143 ACN121999143 ACN 121999143ACN-121999143-A

Abstract

The invention provides an incremental SfM reconstruction method based on relative orientation uncertainty, which relates to the technical field of computer vision, and comprises the steps of obtaining a plurality of images to be reconstructed, calculating relative orientation uncertainty indexes between each pair of images in the plurality of images, wherein the relative orientation uncertainty indexes are calculated based on minimum singular values of a relative orientation coefficient matrix; the method comprises the steps of taking images as nodes, taking a connection relation between the images as edges, giving weight to each edge to construct an undirected graph model, carrying out negative correlation on the weight and relative orientation uncertainty indexes, selecting a node with the highest weight and an adjacent node with the lowest relative orientation uncertainty index of the node to form an initial image pair, calculating an initial relative pose, selecting a node with the lowest relative orientation uncertainty index from alternative nodes connected with a current model, adding the node into the model, calculating the pose of a camera of a newly added image, gradually adding the image until all the images are added, obtaining a reconstruction path with the minimum uncertainty, and further completing model reconstruction.

Inventors

GUO PEIPEI
HONG YONG
WANG MI
ZHANG TINGTING
CAI YULIN

Assignees

湖北珞珈实验室
武汉大学

Dates

Publication Date: 20260508
Application Date: 20260409

Claims (10)

1. The incremental SfM reconstruction method based on the relative orientation uncertainty is characterized by comprising the following steps: acquiring a plurality of images to be reconstructed, and calculating a relative orientation uncertainty index between each pair of images in the plurality of images, wherein the relative orientation uncertainty index is calculated based on the minimum singular value of a relative orientation coefficient matrix; Taking the images as nodes, taking the connection relation between the images as edges, and giving weight to each edge to construct an undirected graph model, wherein the weight is inversely related to the relative orientation uncertainty index; selecting the node with the highest weight and the adjacent node with the lowest relative orientation uncertainty index of the node to form an initial image pair, and resolving the initial relative pose; And selecting a node with the lowest relative orientation uncertainty index from the alternative nodes connected with the current model, adding the node into the model, resolving the camera pose of the newly added image, and sequentially adding the images until all the images are added, so as to obtain a reconstruction path with the minimum uncertainty and further finish the reconstruction of the three-dimensional model.
2. The incremental SfM reconstruction method based on relative orientation uncertainty of claim 1, wherein after each new image is added and its pose and three-dimensional coordinates of the newly added point are restored, a reprojection error of the current model is calculated, and whether the reprojection error is larger than a preset threshold value is judged; If the light beam method adjustment value is larger than the preset threshold value, carrying out light beam method adjustment optimization on the current model; If the light beam method adjustment value is smaller than or equal to the preset threshold value, the light beam method adjustment optimization is not performed; and after all the images are added, performing global beam method adjustment optimization on the whole model.
3. The incremental SfM reconstruction method based on relative orientation uncertainty of claim 1, wherein the calculating of the relative orientation uncertainty step specifically comprises: For each pair of images, extracting homonymous image points, and constructing a relative orientation coefficient matrix based on a coplanar condition, wherein the relative orientation coefficient matrix is constructed based on polar geometric constraint and a coplanar condition equation between the two images; and carrying out singular value decomposition on the relative orientation coefficient matrix to obtain the minimum singular value, and taking the minimum singular value as a measurement value of relative orientation uncertainty between two images.
4. A relative orientation uncertainty-based incremental SfM reconstruction method as in claim 3, wherein said minimum singular value characterizes relative orientation uncertainty, said weight is the inverse of said relative orientation uncertainty indicator, and its calculation formula is: ; Wherein: For the edge weight to be the weight of the edge, Is the minimum singular value of the matrix of relative orientation coefficients.
5. The incremental SfM reconstruction method based on relative orientation uncertainty of claim 1 wherein the node with the highest weight is the node with the highest sum of weights of the connected edges among all edges.
6. The incremental SfM reconstruction method based on relative orientation uncertainty of claim 1, wherein the step of successively adding images specifically comprises: selecting the edge with the lowest relative orientation uncertainty index by taking all the connecting edges of the current model as the reference, and taking all the nodes connected with the edge with the lowest relative orientation uncertainty index as the candidate nodes; calculating the uncertainty sum of the connection of each alternative node and the current model, and selecting the node with the lowest uncertainty sum to add into the model; Wherein the reconstruction path is formed by a sequence of edges and nodes with minimal uncertainty accumulation.
7. The incremental SfM reconstruction method based on relative orientation uncertainty as recited in claim 6, wherein the sum of the uncertainties of the candidate node connection with the current model is the sum of the relative orientation uncertainties of the candidate node connection with all edges of the reconstructed nodes in the current model.
8. An incremental SfM reconstruction system based on relative orientation uncertainty, comprising: The index calculation module is used for acquiring a plurality of images to be reconstructed and calculating relative orientation uncertainty indexes between each pair of images in the plurality of images, wherein the relative orientation uncertainty indexes are calculated based on the minimum singular value of a relative orientation coefficient matrix; The model construction module is used for constructing an undirected graph model by taking the images as nodes and the connection relation between the images as edges and giving weight to each edge, wherein the weight is inversely related to the relative orientation uncertainty index; the initial image pair generating module is used for selecting a node with the highest weight and an adjacent node with the lowest relative orientation uncertainty index of the node to form an initial image pair, and resolving an initial relative pose; The three-dimensional model reconstruction module is used for selecting a node with the lowest relative orientation uncertainty index from the alternative nodes connected with the current model, adding the node into the model, calculating the camera pose of the newly added image, adding the image successively until all the images are added, obtaining a reconstruction path with the minimum uncertainty, and further completing the three-dimensional model reconstruction.
9. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the relative orientation uncertainty based incremental SfM reconstruction method of any one of claims 1 to 7 when the computer program is executed.
10. A non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor, implements the steps of the relative orientation uncertainty based incremental SfM reconstruction method of any one of claims 1 to 7.

Description

Incremental SfM reconstruction method based on relative orientation uncertainty Technical Field The invention relates to the technical field of computer vision, in particular to an incremental SfM reconstruction method based on relative orientation uncertainty. Background The motion restoration structure (SfM) algorithm is a core technology for realizing three-dimensional reconstruction in the fields of computer vision and photogrammetry, wherein incremental SfM is one of the most widely applied strategies. In the scheme, two images are selected from a data set as an initial image pair by taking the overlapping degree as a main basis, the pose of the initial camera and the sparse three-dimensional point cloud are recovered, then a newly added image with the highest overlapping degree with the current model is added in an iterative mode, and the adjustment is optimized by utilizing a beam method until all the images are processed, and finally a complete three-dimensional model is formed. However, the core reconstruction path of the current scheme is severely dependent on the empirical standard of overlapping degree or the number of homonymous points, but lacks the inherent accuracy of relative orientation calculation, namely quantitative evaluation and utilization of uncertainty, and the accuracy is obviously influenced by various key factors such as the length of a base line, the intersection angle, the rough difference of observed values and the like. Therefore, the connection path with the strongest geometric constraint and the most robust solution cannot be guaranteed to be selected only by maximizing the overlapping degree, so that the incremental reconstruction process is extremely sensitive to initial selection, accumulated errors and model drift are extremely easy to introduce, and finally the accuracy and stability of large scene reconstruction are limited. Disclosure of Invention The invention aims to provide an incremental SfM reconstruction method based on relative orientation uncertainty, which aims to solve the problems that the prior scheme core reconstruction path mentioned in the background art is extremely sensitive to initial selection and is extremely easy to introduce accumulated errors and model drift due to the fact that the core reconstruction path is seriously dependent on the overlapping degree or the number of homonymous points. The incremental SfM reconstruction method based on the relative orientation uncertainty comprises the following steps of obtaining a plurality of images to be reconstructed, calculating relative orientation uncertainty indexes between each pair of images in the plurality of images, calculating the relative orientation uncertainty indexes based on minimum singular values of a relative orientation coefficient matrix, taking the images as nodes, constructing an undirected graph model by taking connection relations among the images as edges, giving weights to each edge, wherein the weights are inversely related to the relative orientation uncertainty indexes, selecting a node with the highest weight and an adjacent node with the lowest relative orientation uncertainty index of the node to form an initial pair, calculating an initial relative pose, selecting a node with the lowest relative orientation uncertainty index from alternative nodes connected with a current model, adding the model into the camera pose of a newly-added image, sequentially adding the images until all the images are added, obtaining a reconstruction path with the minimum uncertainty, and further completing the three-dimensional model reconstruction. Optionally, after each image is newly added and the pose and the three-dimensional coordinates of the newly added point are restored, calculating the reprojection error of the current model, judging whether the reprojection error is larger than a preset threshold value, if so, performing beam adjustment optimization on the current model, if not, not performing beam adjustment optimization, and after all images are added, performing global beam adjustment optimization on the whole model. The method comprises the steps of obtaining a relative orientation uncertainty, namely, obtaining a relative orientation coefficient matrix, and obtaining a minimum singular value of the relative orientation uncertainty, wherein the relative orientation coefficient matrix is constructed based on a coplanarity condition, and the relative orientation coefficient matrix is constructed based on polar geometric constraint and a common condition equation between two images. Optionally, the minimum singular value characterizes the relative orientation uncertainty, the weight is the inverse of the relative orientation uncertainty index, and the calculation formula is: Wherein: For the edge weight to be the weight of the edge, Is the minimum singular value of the matrix of relative orientation coefficients. Optionally, the node with the highest weight is the node with t