Non-coding point binocular matching method and system based on multi-level geometric topology consistency

Abstract

The application discloses a non-coding point binocular matching method and system based on multi-level geometric topology consistency, and the method comprises the following steps: respectively extracting the coordinates of non-coding points in a left image and a right image to obtain left and right image feature point sets; constructing an initial candidate matching set based on a geometric constraint of an epipolar line, and dividing the candidate matching set into a unique matching subset and a polysemous matching subset according to the number of candidate matching points; performing multi-scale neighborhood topology consistency screening on the initial candidate matching set, and retaining candidate matching point pairs that meet the topology consistency condition at all scales; establishing a local affine transformation model for the updated unique matching subset, screening geometric anchor point pairs through affine transformation residuals; for the updated polysemous matching subset, estimating a local homography matrix based on geometric anchor point pairs in a local neighborhood, and obtaining a matching point pair set through position prediction and disambiguation matching. The application effectively solves the matching ambiguity problem of non-coding points by combining local geometric guidance with multi-scale topology constraints.

Description

Technical Field

[0001] This invention relates to the field of 3D reconstruction in computer vision, and more particularly to a non-coded point binocular matching method and system based on multi-level geometric topological consistency. Background Technology

[0002] Binocular vision is an important research direction in computer vision and 3D reconstruction. It involves acquiring two images from left and right cameras, calculating disparity using image matching techniques, and then combining this with a camera imaging model to estimate scene depth. The number and accuracy of matching points directly affect the accuracy of 3D measurement and reconstruction; therefore, obtaining a stable and reliable set of matching points in complex scenes has always been a core problem in this field.

[0003] Existing binocular matching methods can be broadly categorized into two types: one is pixel-level matching methods based on grayscale, texture, or descriptors, such as matching based on features like SAD, NCC, SIFT, or ORB. These methods are simple to implement, but are prone to generating a large number of mismatches under conditions of repetitive textures, weak textures, or varying lighting. The other type is geometric consistency methods based on geometric constraints, such as those utilizing epipolar constraints, disparity gradient constraints, or vector angle consistency constraints. These methods can eliminate obvious mismatches to some extent, but since most methods rely only on a single scale or local geometric relationships, many problems still exist.

[0004] Therefore, it is necessary to propose a binocular image matching method to maintain local and global geometric consistency at different scales, thereby significantly reducing the mismatch rate and improving the stability and robustness of the matching point set. Summary of the Invention

[0005] The main objective of this invention is to provide a non-coded point binocular matching method and system based on multi-level geometric topological consistency, which solves the problem that existing binocular matching methods are prone to mismatch and weak matching in complex textures, repetitive structures, and scenes with large scale changes. By performing geometric consistency detection on candidate matching points at multiple scales and fusing the results of different scales, it improves the stability and robustness of matching while ensuring matching accuracy.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a non-coded point binocular matching method based on multi-level geometric topology consistency, comprising the following steps: S1. Simultaneously acquire left and right images of the target scene using a binocular vision acquisition system, and preprocess the images to extract the sub-pixel centroid coordinates of non-coded marker points, thus obtaining the feature point set of the left image. and right image feature point set ; S2. Calculate the fundamental matrix based on the bi-objective parameter determination. For the feature points of the left image For each point in the image, calculate its epipolar line in the right image. By calculating the distance from the feature points in the right image to the epipolar line, construct an initial candidate matching set. Then, divide the candidate matching set into a unique matching subset and a polysemous matching subset based on the uniqueness of the number of candidate matches. S3. Perform multi-scale neighborhood topology consistency screening on the initial candidate matching set, specifically as follows: ① Construct multiple K-nearest neighbor structures with different neighborhood scales in the left and right images respectively; ② Perform consistency calculation on the neighborhood mapping relationship of candidate matching point pairs at different neighborhood scales; ③ Retain the matching point pairs that satisfy the topological consistency condition at multiple scales to obtain the updated unique matching subset and ambiguous matching subset; S4. Establish a local affine geometric model for the updated unique matching subset, and select high-confidence matching points as geometric anchor points by calculating the affine transformation residuals. S5. For the updated polysemous matching subset, the local homography matrix is ​​estimated in the local neighborhood using the geometric anchor points. The theoretical position of the polysemous matching point in the right image is predicted based on the homography matrix. The point with the smallest distance from the predicted position is selected as the final match from the candidate matching point set. S6. Based on the final matching points, construct a Delaunay triangulation network in the left image and map it to the right image. By detecting the topological intersection relationship between corresponding edges in the left and right images, remove mismatched points that disrupt the global topological consistency and obtain the final stable set of matching points.

[0007] Following the above technical solution, step S1 specifically involves: simultaneously acquiring left and right images of the object under test using a calibrated binocular stereo vision sensor; denoising the original image using Gaussian filtering; and extracting non-coded marker regions using Otsu's method for adaptive threshold segmentation. Subsequently, edge detection is performed on the binarized region to extract the contour. Based on the Gaussian-filtered denoised original grayscale image, sub-pixel-level centroid coordinates are calculated within the contour range using the grayscale squared weighted centroid method, generating the left image feature point set P. L With the right image feature point set P R .

[0008] Following the above technical solution, step S2 specifically involves: obtaining the fundamental matrix based on the bi-objective parameter determination. For the left image point set For each point in the right image, calculate its corresponding epipolar equation on the right image; iterate through the point set of the right image. Calculate the vertical Euclidean distance from each point to the polar line, and take the points whose distance is less than the preset tolerance threshold as the initial candidate matching points; count the number of candidate points corresponding to each point in the left figure. If the number is 1, it is assigned to the "unique matching subset"; if the number is greater than 1, it is assigned to the "polysemous matching subset".

[0009] Following the above technical solution, step S3 specifically involves: to eliminate mismatched points that only happen to be valid under epipolar constraints, a multi-scale neighborhood topological consistency screening is performed on the initial candidate matching set; firstly, multiple neighborhood scale parameters are set and K-nearest neighbor structures at different scales are constructed; the K nearest neighbors of each feature point are searched in the feature point set of the left image to form a neighborhood set, and the K nearest neighbors of the corresponding matching point are searched in the feature point set of the right image to form a neighborhood set; then, the mapping consistency between the left and right neighborhoods is calculated based on the candidate matching relationship, and the topological similarity score is obtained by counting the number of point pairs that maintain the matching relationship in the neighborhood; only matching point pairs that meet the similarity threshold condition at all scales are retained, thereby improving the stability of the matching results at different spatial scales.

[0010] Following the above technical solution, step S4 specifically involves: establishing a local affine geometric model for the unique matching subset after multi-scale screening. Specifically, for each matching point, several nearest neighbor matching points are selected, and the local affine transformation matrix from the left image to the right image is solved using the least squares method. The residual between the predicted position after the affine transformation and the actual matching point is calculated. When the residual is less than a preset threshold, the matching point is considered to have high geometric consistency and is marked as a geometric anchor point. If the residual exceeds the threshold, the matching point is considered to be a possible mismatch and is removed, thereby constructing a set of reliable high-confidence matching points as the basis for subsequent matching guidance.

[0011] Following the above technical solution, step S5 specifically involves: for cases where multiple candidate matching points exist in the polysemous matching subset, a local homography geometric model is established in the local neighborhood using the geometric anchor points obtained in step S4. The homography matrix is ​​estimated using the least squares method by selecting a certain number of neighboring anchor point pairs. Subsequently, the homography matrix is ​​used to project and predict the polysemous feature points in the left image, calculate their theoretical positions in the right image, and calculate the Euclidean distance between the predicted position and all candidate matching points. Finally, the candidate point with the smallest distance from the predicted position and satisfying the epipolar constraint condition is selected as the final match for the feature point, thereby effectively eliminating the ambiguity in matching caused by repetitive structures or common epipolar phenomena.

[0012] Following the above technical solution, step S6 specifically involves: based on the final obtained stable matching point set, constructing a Delaunay triangulation network in the feature point set of the left image to establish topological connections between points, and mapping the triangulation edges in the left image to the right image to form corresponding edge sets according to the matching relationships; then traversing all corresponding edge pairs, using a line segment intersection detection algorithm to determine whether the corresponding edges in the right image intersect, while the corresponding edges in the left image do not intersect; if such topological anomalies are detected, it is considered that the relevant matching points have disrupted the consistency of the overall geometric structure, and vertices with lower confidence are preferentially removed according to the matching confidence, and the detection is repeated until all topological conflicts are eliminated, thereby obtaining the final stable matching point set that satisfies global topological consistency.

[0013] In a second aspect, the present invention provides a non-coded point binocular matching system based on multi-level geometric topology consistency, for implementing the method described in the first aspect, the system comprising: The image acquisition and feature extraction module uses a binocular camera to acquire left and right images of the object under test. First, the images are filtered, denoised, and binarized to separate non-coded point regions. Then, the edge contours of non-coded marker points are extracted, and the gray-level squared weighted centroid method is used to calculate high-precision sub-pixel centroid coordinates, thereby obtaining the basic feature point set of the left and right images. The epipolar constraint and classification module establishes epipolar geometric constraints based on bi-objective calibration parameters, calculates the vertical distance from feature points in the right image to the corresponding epipolar line in the left image, and filters out initial candidate matches that meet the threshold. Based on the uniqueness of the number of candidate points, the initial set is divided into an unambiguous unique matching subset and a polysemous matching subset with multiple points competing, laying the foundation for subsequent splitting processing. The multi-scale consistency screening module performs dynamic multi-scale checks on the initial candidate matching set, constructs K-nearest neighbor graphs in three different neighborhood ranges (micro, meso, and macro), calculates the neighborhood topological similarity of matching point pairs at different scales, retains only matching points that show high robustness at all scales, and removes noise interference that only occasionally meets the conditions at a single scale. The anchor point guidance and disambiguation module adopts a hierarchical guidance strategy. First, it uses a local affine transformation model to perform geometric verification on the unique matching subset and establish high-confidence geometric anchor points. Then, it uses these anchor points to construct a local homography matrix to accurately predict the spatial location of points in the ambiguous matching subset. By calculating the distance deviation between the candidate points and the theoretical predicted coordinates, it eliminates ambiguity and locks in the best match. The global topology verification module constructs a Delaunay triangulation network based on the retained matching points, maps the grid connection relationship of the left image to the right image, and detects whether there are abnormal situations where corresponding edges of the left and right sides intersect. By removing mismatched points that cause the global topology structure to flip or break, it ensures that the final output set of matching points maintains continuity and consistency in physical space. The inter-frame model construction module, based on the set of rotation and translation transformation matrices, applies them to the 3D information of non-coded points in different frames to construct a 3D spatial model of the non-coded points of the object under test.

[0014] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: This invention lays a solid foundation for high-precision initial matching through sub-pixel centroid extraction and epipolar constraint classification. It innovatively introduces a dynamic multi-scale neighborhood consistency screening mechanism, effectively overcoming local topological instability caused by viewpoint scaling or depth-of-field changes at a single scale, and significantly reducing the initial mismatch rate. Addressing the ambiguity problem in matching within common epipolar regions, a local homography guidance strategy based on geometric anchor points is established. This strategy utilizes deterministic information to accurately deduce the positions of ambiguous points, significantly outperforming traditional sequential constraints under complex occlusion or large viewpoint rotation. Furthermore, by combining Delaunay triangulation global topological cross-detection, residual mismatch points causing structural flipping are eliminated from the perspective of physical spatial continuity. The comprehensive application of these technologies significantly enhances the system's adaptability in unstructured environments, ensuring high accuracy and robustness of non-coded point 3D reconstruction, and providing reliable technical support for handheld binocular multi-line structured light 3D reconstruction. Attached Figure Description

[0015] Figure 1 This is a flowchart of a non-coded point binocular matching method based on multi-level geometric topology consistency according to an embodiment of the present invention; Figure 2 This is an image captured by a binocular camera according to an embodiment of the present invention; Figure 3 This is a diagram showing the Gaussian filtering denoising result according to an embodiment of the present invention; Figure 4 This is an image showing the adaptive binarization result of applying Otsu's method according to an embodiment of the present invention; Figure 5 This is an edge map obtained using the Canny operator according to an embodiment of the present invention; Figure 6 This is an image showing the results of non-coded point edge detection and sub-pixel centroid extraction according to an embodiment of the present invention; Figure 7 This is an initial candidate matching line diagram based on epipolar geometry and row alignment constraints according to an embodiment of the present invention; Figure 8This is a matching result diagram based on multi-scale neighborhood topology consistency screening according to an embodiment of the present invention; Figure 9 This is a matching result diagram after disambiguation guided by local homography according to an embodiment of the present invention; Figure 10 This is a diagram showing the final high-precision two-dimensional matching result of corresponding points after global topology cross detection according to an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments provided by this invention without inventive effort are within the scope of protection of this invention.

[0017] Obviously, the accompanying drawings described below are merely some examples or embodiments of the present invention. Those skilled in the art can apply the present invention to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this invention, modifications to design, manufacturing, or production based on the technical content disclosed in this invention are merely conventional technical means and should not be construed as insufficient disclosure of the present invention.

[0018] In this invention, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this invention may be combined with other embodiments without conflict.

[0019] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "a," "an," "an," "the," and similar words used in this invention do not indicate quantity limitation and may indicate singular or plural. The terms "comprising," "including," "having," and any variations thereof used in this invention are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms "connected," "linked," "coupled," and similar words used in this invention are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "A plurality" used in this invention refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships may exist; for example, "A and / or B" can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects have an "or" relationship. The terms "first," "second," and "third" used in this invention are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0020] This invention provides a binocular matching method and system for non-coded points based on multi-level geometric topological consistency, belonging to the field of computer vision 3D measurement technology. The method first uses a binocular camera to acquire images and calculate the sub-pixel centroid coordinates of non-coded points. An initial candidate matching set is constructed based on epipolar geometric constraints, and then divided into unique matching subsets and ambiguous matching subsets according to the uniqueness of the number of candidate matches. Subsequently, K-nearest neighbor graphs are constructed under different K values ​​to dynamically screen the initial candidate matching set for neighborhood consistency at multiple scales, retaining matching points with stable topological structures at all scales. Based on this, a hierarchical guidance strategy is adopted: first, a local affine transformation model is used to verify the unique matching subset to establish high-confidence geometric anchor points; then, a local homography matrix is ​​constructed using the anchor points to perform accurate spatial location prediction and disambiguation matching for the ambiguous matching subsets. Finally, a Delaunay triangulation network is constructed based on the retained point set, and residual mismatched points that disrupt global planarity are eliminated by detecting topological intersection anomalies on corresponding edges of the left and right images. This invention effectively solves the matching ambiguity problem of non-coded points in complex scenes by combining local geometric guidance with global topological constraints, and significantly improves the accuracy and robustness of 3D reconstruction.

[0021] Example 1: like Figure 1 As shown, the non-coded point binocular matching method based on multi-level geometric topology consistency according to an embodiment of the present invention includes the following steps: S1. Simultaneously acquire left and right images of the target scene using a binocular vision acquisition system, and preprocess the images to extract the sub-pixel centroid coordinates of non-coded marker points, thus obtaining the feature point set of the left image. and right image feature point set ; S2. Calculate the fundamental matrix based on the bi-objective parameter determination. For the feature points of the left image For each point in the image, calculate its epipolar line in the right image. By calculating the distance from the feature points in the right image to the epipolar line, construct an initial candidate matching set. Then, divide the candidate matching set into a unique matching subset and a polysemous matching subset based on the uniqueness of the number of candidate matches. S3. Perform multi-scale neighborhood topology consistency screening on the initial candidate matching set, specifically as follows: ① Construct multiple K-nearest neighbor structures with different neighborhood scales in the left and right images respectively; ② Perform consistency calculation on the neighborhood mapping relationship of candidate matching point pairs at different neighborhood scales; ③ Retain the matching point pairs that satisfy the topological consistency condition at multiple scales to obtain the updated unique matching subset and ambiguous matching subset. S4. Establish a local affine geometric model for the updated unique matching subset, and select high-confidence matching points as geometric anchor points by calculating the affine transformation residuals. S5. For the updated polysemous matching subset, the local homography matrix is ​​estimated in the local neighborhood using the geometric anchor points. The theoretical position of the polysemous matching point in the right image is predicted based on the homography matrix, and the point with the smallest distance from the predicted position is selected as the final match from the candidate point set. S6. Based on the final matching points, construct a Delaunay triangulation network in the left image and map it to the right image. By detecting the topological intersection relationship between corresponding edges in the left and right images, remove mismatched points that disrupt the global topological consistency and obtain the final stable set of matching points.

[0022] The non-coded point binocular matching method based on multi-level geometric topology consistency in this invention is built on a Windows 10 platform using C++ and QT. It is understood that the algorithm of this invention can also be ported to other programming languages ​​or hardware devices.

[0023] Example 2: The non-coded point binocular matching method based on multi-level geometric topology consistency in this embodiment is the same as that in Example 1, except that each step is applied to a specific instance, specifically including the following steps: In step S1, image preprocessing and sub-pixel feature extraction are performed by simultaneously capturing images of the object under test using a calibrated binocular camera to obtain the left image. And the right image ， like Figure 2 As shown. Gaussian filtering is applied to the image for denoising, and the denoising result is as follows. Figure 3 As shown. Binarization is performed using the Otsu method to separate non-coded point regions. The binarization result is shown below. Figure 4 As shown in the figure. The Canny operator is used to extract the edge contours of non-coded points, and the resulting edge map is shown in the figure. Figure 5 As shown. To improve matching accuracy, instead of directly using the binarized center, the original grayscale image is used (preferably the image after Gaussian filtering and denoising, and secondarily the original left image). And the right image The sub-pixel coordinates of feature points are calculated using the gray-level squared weighted centroid method.

[0024] Specifically, the left image of the target scene is first acquired simultaneously using a binocular vision acquisition system. And the right image To reduce the impact of noise on feature extraction, the original image is processed using a Gaussian filter: ; in, The original image grayscale values. This is the Gaussian kernel function.

[0025] Subsequently, Otsu's method was used for adaptive threshold segmentation to obtain a binary image. ; in, The optimal threshold is calculated using the Otsu method. For binary images in The pixel value at the location.

[0026] Contour extraction is performed on the binary image to obtain the region of each non-coded marker point. Ω In the original grayscale image, the grayscale squared weighted centroid of each non-coded marker region is calculated, thereby obtaining the sub-pixel level feature point coordinates of each non-coded marker: ; This yields the feature point sets for the left and right images: ; in: ; In the formula, and These are the sub-pixel level feature point coordinates of the non-coded marker points. and Let the feature point set of the left image and the feature point set of the right image be... For the first One left image feature point, For the first One left image feature point, The number of feature points in the left image. This represents the number of feature points in the right image.

[0027] The final extraction result is as follows Figure 6 As shown.

[0028] In step S2, a classification strategy based on epipolar constraints is used to perform extreme constraint screening on feature points using the calibration parameters of a binocular camera, and the point set is divided into two categories based on the uniqueness of the match for subsequent calculations.

[0029] Specifically, the epipolar line is first calculated for the first epipolar line in the left image. Feature points Its corresponding polar line in the right image for: ; In the formula, The base matrix is ​​calculated based on the camera's intrinsic and extrinsic parameters and represents the spatial positional relationship between the two cameras.

[0030] Next, distance determination is performed, calculating all feature points in the right image. to the poles Normalized Euclidean distance : ; Set limit constraint threshold (This example takes) (pixels), if Then it is believed yes Candidate matching points.

[0031] Finally, the dataset is partitioned, and the points in the left image are counted. The corresponding number of candidate matching points .like That is, if there is only one point in the right figure that satisfies the epipolar constraint, then the matching pair... Included in the "unique matching subset"; if That is, there exist multiple points in the right figure that satisfy the epipolar constraint (this is ambiguous), and this group of points... Included in the "polysemous matching subset", the initial matching result is as follows: Figure 7 As shown.

[0032] In step S3, the multi-scale neighborhood topology consistency screening is performed. Since step S2 only utilizes epipolar geometric constraints to construct the candidate matching set, some mismatched points may still exist. For example, multiple candidate points may simultaneously satisfy the epipolar distance constraint due to dense distribution or repetitive structures of non-coded points. Therefore, it is necessary to further utilize the spatial neighborhood structure relationship between feature points to screen candidate matches. This step establishes a multi-scale neighborhood topology structure to evaluate the neighborhood consistency of candidate matching point pairs at different spatial scales, thereby eliminating erroneous matches that disrupt the local topology structure.

[0033] This invention first determines the left K-nearest neighbor (KNN) of the left image feature point in the candidate matching point pair and the right KNN of the right image feature point in the candidate matching point pair. Then, for each feature point in the left KNN, it determines whether at least one of the candidate matching points of the feature point is located in the right KNN. If so, it is counted as a left image feature point that maintains a matching relationship. The number of left image feature points that maintain a matching relationship is counted, and the ratio of the number of left image feature points to the total number of feature points in the left KNN at the neighborhood scale is calculated as the topological similarity score.

[0034] Specifically, K-nearest neighbor structures are constructed in the feature point sets of the left and right images, respectively. Let the neighborhood scale set be: These correspond to the neighborhood structures at the micro, meso, and macro scales, respectively.

[0035] For feature points in the left image In the set of feature points Searching for it Calculate the nearest neighbor points and construct a neighborhood set: , .

[0036] For candidate matching points in the right image Similarly, construct the corresponding neighborhood set: .

[0037] After obtaining the neighborhood sets of the left and right images, a neighborhood mapping is established through candidate matching relationships, and a neighborhood topology consistency score is calculated. The neighborhood topology consistency score is defined as: ; in: ; In the formula, Point The candidate matching points, where the numerator represents the number of points that maintain neighborhood correspondence, and the denominator represents the number of points that maintain neighborhood correspondence. This represents the total number of neighborhood points at this scale.

[0038] After completing the above steps, perform robustness screening and set a consistency threshold. If and only if the matching point pairs satisfy the following conditions at all three scales. Only then will it be retained. That is: ; This step effectively removes false matches that appear correct in the local view but do not match the topology at a large scale. The resulting matching image after removal is shown below. Figure 8 As shown.

[0039] It should be noted that after multi-scale neighborhood topological consistency screening, some candidate matching point pairs in the initial candidate matching set are eliminated because they do not meet the topological consistency condition, thus obtaining the updated unique matching subset and polysemous matching subset. Specifically, matching point pairs in the original unique matching subset may be deleted because they fail the topological consistency test at a certain scale; candidate matching point pairs in the original polysemous matching subset may also become unique matching point pairs due to the elimination of some candidate points and be included in the updated unique matching subset, or all candidate points may be eliminated and deleted entirely. Furthermore, in subsequent steps of this invention, both the unique matching subset and the polysemous matching subset refer to the subsets updated after multi-scale neighborhood topological consistency screening. That is, a local affine transformation model is established for the updated unique matching subset to screen geometric anchor point pairs, and disambiguation matching based on the local homography matrix is ​​performed on the updated polysemous matching subset.

[0040] In step S4, local affine model screening and geometric anchor point establishment are performed. After multi-scale topological screening in step S3, the candidate matching point set already satisfies the basic neighborhood structure consistency. However, in local regions, there may still be geometrically distorted matches caused by noise or uneven distribution of feature points. Therefore, this step further utilizes a local geometric transformation model to finely screen the matching points and selects matching points with higher geometric stability as geometric anchor points to provide a reliable reference for subsequent ambiguous matching elimination.

[0041] Specifically, within the uniquely matched subset, several neighboring matching points are selected centered on a given matching point. The local affine transformation model from the left image to the right image is estimated using the least squares method. The local affine transformation relationship can be expressed as: ; in, This is the local affine transformation matrix.

[0042] The affine transformation model can be used to predict the positions of feature points in the left image within the right image, and the residual error between the predicted positions and the actual matching points can be calculated. ; in This represents the Euclidean norm.

[0043] When the residual satisfies the following conditions < When a matching point satisfies the local geometric consistency constraint, it is marked as a geometric anchor point. All matching points that satisfy the condition constitute the set of geometric anchor points. The matching points in this set exhibit high geometric stability in spatial distribution and can serve as important reference points for subsequent polysemous matching prediction.

[0044] It should be noted that when establishing the local affine transformation model, the selected neighboring matching point pairs all come from the updated unique matching subset. That is, all the points participating in the estimation of the affine transformation matrix are reliable matching point pairs with confirmed one-to-one correspondences. Furthermore, the estimation direction of the local affine transformation matrix can be from the left image to the right image, or from the right image to the left image. Correspondingly, the positions of feature points in the left image can be predicted in the right image and the residuals calculated, or the positions of feature points in the right image can be predicted in the left image and the residuals calculated. Either approach is acceptable, and the purpose is to verify the geometric consistency of the matching point pairs to select geometric anchor point pairs.

[0045] In step S5, ambiguous matching is eliminated based on homography prediction. In real-world scenarios, due to the lack of unique identifiers among non-coded marker points, some feature points may correspond to multiple candidate matching points under epipolar constraints, thus forming ambiguous matching relationships. To further determine the correct matching points, this step uses the set of geometric anchor points obtained in step S4 to establish a local homography transformation model, and eliminates ambiguous matching through geometric prediction.

[0046] Specifically, several geometric anchor points are selected around the polysemous matching points, and the local homography matrix is ​​estimated using the least squares method with these anchor point pairs: ; The homography matrix describes the projection transformation relationship between the left and right images in a local planar region.

[0047] For the polysemous points in the left image The theoretical position of the image in the right image is predicted using the homography matrix: ; in: ; The predicted coordinates are obtained by normalizing the homogeneous coordinates: ; Then, the Euclidean distance between the predicted location and all candidate matching points is calculated: ; Finally, the candidate point with the smallest distance is selected as the final matching point: ; And this distance is required to satisfy the following constraints: ; In the formula, The second polar constraint threshold is less than the limit constraint threshold. .

[0048] This predictive matching process effectively eliminates matching ambiguity caused by repetitive structures or dense point distributions, thereby further improving matching accuracy. The matching results are as follows: Figure 9 As shown.

[0049] In step S6, global topology consistency optimization based on Delaunay triangulation, after eliminating ambiguous matches in step S5, has yielded a relatively reliable set of matching point pairs. However, in complex scenarios, a small number of mismatched points may still exist. Although these mismatched points satisfy local geometric constraints, they may disrupt the topological relationships between matching points in the overall spatial structure. Therefore, this step further optimizes the matching results by constructing a global topology structure and detecting topological conflicts to obtain a stable and reliable final set of matching points.

[0050] Specifically, a Delaunay triangulation network is first constructed using the set of matching points in the left image. Delaunay triangulation establishes topological connections between points while ensuring that the triangle shape is as close as possible to an equilateral triangle. This better reflects the spatial distribution structure of the point set. Let the set of matching points in the left image be: ; The topology is obtained through Delaunay triangulation: ; in, For the set of vertices, It is the set of edges of a triangular mesh.

[0051] Based on the matching relationship, the edges in the left image are mapped to the right image to obtain the corresponding set of edges: ; After obtaining the corresponding edge sets of the left and right images, a line segment intersection detection method is used to determine whether the topological structure remains consistent. The line segment intersection determination function is defined as follows: ; For any two edges , and its corresponding edge in the right image , If the following conditions are met: ; like: ; This indicates that the matching point pair has disrupted the original topological relationship between the left and right images. That is, the edges in the left image do not intersect, while the corresponding edges in the right image intersect. This type of situation is usually caused by mismatched points.

[0052] To eliminate this type of topological conflict, this embodiment further introduces a matching confidence evaluation function to filter matching points involved in the conflict. The matching point confidence can be defined as: ; in, The affine residual calculated in step S4; This indicates the reliability of the matching point.

[0053] When a topological conflict is detected, matching points with lower confidence are removed first, and the topological relationships between the remaining matching points are re-examined. This process is iterated until all edge pairs satisfy the topological consistency condition.

[0054] Finally, a stable set of matching points that satisfies global topological consistency is obtained: ; This global topology optimization process effectively eliminates a small number of residual mismatched points, ensuring that the matching results simultaneously satisfy epipolar geometric constraints, local geometric consistency, and global topological consistency. This significantly improves the stability and accuracy of binocular matching, as shown in the matching results below. Figure 10 As shown.

[0055] The final set of points selected is the final stable set of matching points, which is output for subsequent 3D coordinate reconstruction.

[0056] Example 3: This embodiment is a non-coded point binocular matching system based on multi-level geometric topology consistency, used to implement the non-coded point binocular matching method based on multi-level geometric topology consistency in the above method embodiment, including: The image acquisition and feature extraction module uses a binocular camera to acquire left and right images of the object under test. First, the images are filtered, denoised, and binarized to separate non-coded point regions. Then, the edge contours of non-coded marker points are extracted, and the gray-level squared weighted centroid method is used to calculate high-precision sub-pixel centroid coordinates, thereby obtaining the basic feature point set of the left and right images. The epipolar constraint and classification module establishes epipolar geometric constraints based on bi-objective calibration parameters, calculates the vertical distance from feature points in the right image to the corresponding epipolar line in the left image, and filters out initial candidate matches that meet the threshold. Based on the uniqueness of the number of candidate points, the initial set is divided into an unambiguous unique matching subset and a polysemous matching subset with multiple points competing, laying the foundation for subsequent splitting processing. The multi-scale consistency screening module performs dynamic multi-scale checks on the initial candidate matching set, constructs K-nearest neighbor graphs in three different neighborhood ranges (micro, meso, and macro), calculates the neighborhood topological similarity of matching point pairs at different scales, retains only matching points that show high robustness at all scales, and removes noise interference that only occasionally meets the conditions at a single scale. The anchor point guidance and disambiguation module adopts a hierarchical guidance strategy. First, it uses a local affine transformation model to perform geometric verification on the unique matching subset and establish high-confidence geometric anchor points. Then, it uses these anchor points to construct a local homography matrix to accurately predict the spatial location of points in the ambiguous matching subset. By calculating the distance deviation between the candidate points and the theoretical predicted coordinates, it eliminates ambiguity and locks in the best match. The global topology verification module constructs a Delaunay triangulation network based on the retained matching points, maps the grid connection relationship of the left image to the right image, and detects whether there are abnormal situations where corresponding edges of the left and right sides intersect. By removing mismatched points that cause the global topology structure to flip or break, it ensures that the final output set of matching points maintains continuity and consistency in physical space. The inter-frame model construction module, based on the set of rotation and translation transformation matrices, applies them to the 3D information of non-coded points in different frames to construct a 3D spatial model of the non-coded points of the object under test.

[0057] Each module is mainly used to implement the various steps of the method embodiment, which will not be elaborated here.

[0058] It should be noted that, depending on the implementation needs, the various steps / components described in this invention can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0059] Those skilled in the art will readily understand that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A non-coded point binocular matching method based on multi-level geometric topological consistency, characterized in that, The method includes: The coordinates of non-coded points in the left and right images are extracted respectively to obtain the feature point set of the left image and the feature point set of the right image; Determine the epipolar line corresponding to the feature point in the left image in the right image, calculate the distance from each feature point in the right image to the epipolar line, and take the right image feature point whose distance satisfies the epipolar line constraint threshold as the candidate matching point of the left image feature point; construct an initial candidate matching set based on all left image feature points and the candidate matching points corresponding to each left image feature point, and divide the initial candidate matching set into a unique matching subset and a polysemous matching subset according to the uniqueness of the number of candidate matching points; Multi-scale neighborhood topology consistency screening is performed on the initial candidate matching set: multiple K-nearest neighbor structures with different neighborhood scales are constructed in the left and right images respectively, the neighborhood mapping consistency of each candidate matching point pair at different scales is evaluated, and only candidate matching point pairs that satisfy the topology consistency condition at all scales are retained to obtain the updated unique matching subset and ambiguous matching subset. A local affine transformation model is established for the updated unique matching subset. By calculating the affine transformation residuals of feature points in the left or right image, high-confidence candidate matching point pairs are selected as geometric anchor point pairs. For each left image feature point in the updated polysemous matching subset, the local homography matrix is ​​estimated based on the geometric anchor point pairs in the local neighborhood. The position of the left image feature point in the right image is predicted using the homography matrix. The point with the smallest distance from the predicted position is selected from the candidate matching points corresponding to the left image feature point as the final matching point, and finally the set of matching point pairs between the left and right images is obtained. The step of performing multi-scale neighborhood topology consistency screening on the initial candidate matching set includes: Set multiple different neighborhood scale parameters; For each candidate matching point pair in the initial candidate matching set, at each neighborhood scale, K-nearest neighbors are constructed in the left image feature point set and the right image feature point set, respectively. By counting the number of left image feature points that maintain a matching relationship in the left and right K-nearest neighbors, the topological similarity score of the candidate matching point pair at that neighborhood scale is obtained. If a candidate matching point pair satisfies the similarity threshold condition in terms of topological similarity scores at all neighborhood scales, it is retained; otherwise, it is discarded.

2. The non-coded point binocular matching method based on multi-level geometric topological consistency according to claim 1, characterized in that, The step of extracting the coordinates of non-coded points in the left and right images respectively includes: The left and right images are filtered, denoised, and binarized to extract non-coded point regions; Based on the non-coded point region, extract the edge contour of the non-coded points; On the filtered and denoised grayscale image, the sub-pixel level coordinates are calculated for the region within the edge contour using the grayscale square weighted centroid method.

3. The non-coded point binocular matching method based on multi-level geometric topological consistency according to claim 1, characterized in that, The step of using right image feature points whose distance satisfies the epipolar constraint threshold as candidate matching points for the left image feature points includes: Calculate the fundamental matrix based on the bi-objective parameter determination; Based on the fundamental matrix and the coordinates of the feature point in the left image, determine the epipolar line corresponding to the feature point in the left image in the right image; Calculate the distance from each feature point in the right image to the epipolar line; Right image feature points whose distance is less than or equal to the epipolar constraint threshold are used as candidate matching points for the left image feature points.

4. The non-coded point binocular matching method based on multi-level geometric topological consistency according to claim 1, characterized in that, The step of dividing the initial candidate matching set into a unique matching subset and a polysemous matching subset based on the uniqueness of the number of candidate matching points includes: If the number of candidate matching points is 1, then the candidate matching point pair is assigned to the unique matching subset; If the number of candidate matching points is greater than 1, then all candidate matching point pairs corresponding to the feature points of the left image are included in the polysemous matching subset.

5. The non-coded point binocular matching method based on multi-level geometric topological consistency according to claim 1, characterized in that, The topological similarity score is: Determine the left K-nearest neighbor region of the feature point in the left image of the candidate matching point pair, and the right K-nearest neighbor region of the feature point in the right image of the candidate matching point pair; For each feature point in the left K-nearest neighbor region, determine whether there is at least one candidate matching point of the feature point located in the right K-nearest neighbor region. If so, it is counted as a left graph feature point that maintains a matching relationship. The number of feature points in the left image that maintain a matching relationship is counted, and the ratio of the number of feature points in the left image to the total number of feature points in the left K nearest neighbors at the same neighborhood scale is calculated as the topological similarity score. Candidate matching point pairs are retained if their topological similarity scores at all neighborhood scales are higher than the same preset consistency threshold; otherwise, they are discarded.

6. The non-coded point binocular matching method based on multi-level geometric topological consistency according to claim 1, characterized in that, The step of establishing a local affine transformation model for the updated unique matching subset, and selecting high-confidence candidate matching point pairs as geometric anchor point pairs by calculating the affine transformation residuals of feature points in the left or right image, includes: For each candidate matching point pair in the updated unique matching subset, select several neighboring matching point pairs in the updated unique matching subset; Estimate the local affine transformation matrix using selected nearest matching point pairs; Using the local affine transformation matrix, the position of one feature point in the candidate matching point pair in the image of the other feature point is predicted, and the residual between the predicted position and the actual position of the other feature point is calculated. Candidate matching point pairs whose residuals are less than a preset affine residual threshold are marked as the geometric anchor point pairs.

7. The non-coded point binocular matching method based on multi-level geometric topological consistency according to claim 1, characterized in that, The method also includes: A Delaunay triangulation network is constructed based on the set of matching point pairs. Mismatched point pairs are eliminated through global topological consistency, resulting in the final set of matching point pairs, as follows: Based on the set of matching point pairs, a Delaunay triangulation network is constructed using the matching points in the left image; Based on the matching relationship, the edges of the Delaunay triangulation network in the left image are mapped to the right image to obtain the corresponding edge set in the right image; Detect whether there are topologically anomalous edge pairs that do not intersect in the left image but whose corresponding edges intersect in the right image; If such a problem exists, for the matching point pairs that cause topological anomalies, compare the matching confidence of each matching point pair, prioritize removing matching point pairs with lower confidence, and re-examine the topological relationships between the remaining matching points until all edge pairs satisfy topological consistency, and output the final set of matching point pairs.

8. The non-coded point binocular matching method based on multi-level geometric topological consistency according to claim 7, characterized in that, The matching confidence level is defined as follows: ,in, Let be the matching confidence score of the i-th pair of matching points. The affine residual is calculated for the matching point pair under the local affine transformation model.

9. A non-coded point binocular matching system based on multi-level geometric topological consistency, characterized in that, For implementing the method as described in any one of claims 1 to 8, the system comprises: The image acquisition and feature extraction module is used to acquire the left and right images, and extract the coordinates of non-coded points in the left and right images respectively, so as to obtain the feature point set of the left image and the feature point set of the right image. The epipolar constraint and classification module is used to determine the epipolar line corresponding to the feature points of the left image in the right image, calculate the distance from each feature point of the right image to the epipolar line, and take the right image feature points whose distances meet the epipolar constraint threshold as candidate matching points of the left image feature points; based on all left image feature points and the candidate matching points corresponding to each left image feature point, an initial candidate matching set is constructed, and the initial candidate matching set is divided into a unique matching subset and a polysemous matching subset according to the uniqueness of the number of candidate matching points; The multi-scale consistency screening module is used to perform multi-scale neighborhood topological consistency screening on the initial candidate matching set: multiple K-nearest neighbor structures with different neighborhood scales are constructed in the left and right images respectively, the neighborhood mapping consistency of each candidate matching point pair at different scales is evaluated, and only candidate matching point pairs that satisfy the topological consistency condition at all scales are retained to obtain the updated unique matching subset and ambiguous matching subset. The anchor point filtering module is used to establish a local affine transformation model for the updated unique matching subset. By calculating the affine transformation residuals of feature points in the left or right image, high-confidence candidate matching point pairs are selected as geometric anchor point pairs. The disambiguation matching module is used to estimate the local homography matrix based on the geometric anchor point pairs in the local neighborhood for each left image feature point in the updated polysemous matching subset, use the homography matrix to predict the position of the left image feature point in the right image, and select the point with the smallest distance from the predicted position from the candidate matching points corresponding to the left image feature point as the final matching point, and finally obtain the set of matching point pairs between the left and right images. The global topology verification module is used to construct the Delaunay triangulation network based on the set of matching point pairs. It eliminates mismatched point pairs through global topology consistency to obtain the final set of matching point pairs.

Images