A method for initial screening of heterogeneous scene matching point features fused with navigation information
By fusing navigation information, image coordinates are corrected using attitude and relative height, a first-order gradient map is calculated, and weighted principal direction difference filtering is performed. This solves the problem of low coarse matching accuracy in heterogeneous scene matching and achieves efficient image matching results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIAN MODERN CONTROL TECH RES INST
- Filing Date
- 2025-05-27
- Publication Date
- 2026-07-24
AI Technical Summary
In existing technologies, the proportion of correctly matched point pairs in the coarse matching results of point features in heterogeneous scene matching is low, which leads to a decrease in the matching accuracy and an increase in the time consumption of the fine matching process.
By employing a method that integrates navigation information, the image coordinate system is corrected using attitude and relative altitude information. The first-order gradient magnitude and direction map are calculated, and the weighted principal direction difference and matching effect index are used for initial screening of coarse matching point pairs, thereby improving matching accuracy and reducing computational complexity.
It improves the matching accuracy of the coarse matching process by 42%, reduces the time consumption of the fine matching process by 14%~33%, is applicable to all point features, and improves the accuracy of heterogeneous image matching.
Smart Images
Figure CN120599298B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of scene matching technology, and specifically to a method for initial screening of heterogeneous scene matching point features that integrates navigation information. Background Technology
[0002] Currently, point features are commonly used in traditional heterogeneous image matching. This field typically employs a three-step process: point feature extraction, coarse matching, and fine matching. First, point features are extracted, then modality-independent descriptors are extracted from these features. Next, a brute-force matching method is used to select the closest point to each descriptor as the matching point, completing coarse matching. Finally, fine matching is performed using methods such as RANSAC (Random Sample Consensus) and FSC (Fast Sample Consensus) based on the coarse matching results. However, due to the significant differences in descriptor information between heterogeneous images, the proportion of correctly matched point pairs in the coarse matching results is usually low. This leads to a decrease in the matching accuracy of the subsequent fine matching process. Furthermore, the time consumption of the fine matching process, based on RANSAC and FSC, which rely on iterative convergence, is significantly increased in order to select correct matching point pairs from the coarse matching results with a low proportion of correct pairings. Summary of the Invention
[0003] The purpose of this invention is to provide a method for initial screening of heterogeneous scene matching point features that integrates navigation information, thereby improving the positioning accuracy of terrain matching navigation methods while also having high computational efficiency.
[0004] To achieve the above objectives, the present invention employs the following technical solution: A method for initial screening of heterogeneous scene matching point features by fusing navigation information includes: Obtain the real-time image, and perform coarse matching of feature points between the real-time image and the baseline image to obtain the feature point sets of the real-time image and the baseline image. The real-time map is corrected for geographic information to make it consistent with the reference map; for the corrected real-time map and the reference map, the corresponding first-order gradient magnitude map is calculated, and the corresponding first-order gradient direction map is calculated. Pixel-by-pixel mapping is performed on the first-order gradient direction maps of the real-time map and the baseline map to obtain the corresponding partition maps; For the real-time image, the radius is determined based on preset radius parameters and radius coefficients. For the position coordinates of feature points on the corrected real-time image, the position coordinates that are less than the radius from the feature point are traversed, and a histogram is constructed. The principal direction of each feature point is determined based on the histogram and recorded in the first region principal direction array. The radius scaling factor is adjusted twice to obtain the second and third region principal direction arrays. Based on the three region principal direction arrays, the unique principal direction array of feature points in the real-time image is calculated. The same method is used to determine the unique principal direction array of the reference image. The absolute value of the difference between the principal directions is determined by using the unique principal direction arrays of the real-time graph and the baseline graph; Construct an index array by using the absolute value of the difference in the main direction and a preset matching effect index, and sort it. Select the corresponding feature point pair according to the required number of feature point pairs.
[0005] Furthermore, pixel-by-pixel mapping is performed on the first-order gradient direction maps of the real-time map and the baseline map to obtain the corresponding partition maps, including: For the corrected real-time graph Baseline diagram Calculate the first-order gradient pattern , ; For two first-order gradient patterns and Perform pixel-by-pixel mapping and calculate the partition map respectively. ; one of the first-order gradient patterns Mapping yields a partitioned map The process is as follows: For the first-order gradient pattern Each value in the parameter is multiplied O The result is then divided by 360 and rounded; among which, O This is an adjustable threshold.
[0006] Furthermore, for real-time graphs, the radius is determined based on preset radius parameters and radius coefficients, including: Set radius parameter and radius coefficient Then the radius .
[0007] Further, traversing the position coordinates less than the radius from the feature point and constructing a histogram includes: Initialize histogram The q element The histogram is 1* O An array of sizes;
[0008] in, Represents position coordinates In the partitioned graph corresponding to the real-time graph The position above, Indicates to Perform statistical histogram operations. Represents position coordinates The first-order gradient magnitude map corresponding to the real-time graph The position above, Histogram middle The element corresponding to the position.
[0009] Furthermore, the principal direction of each feature point is determined based on the histogram and recorded in the first region principal direction array, including: Find the histogram The indices of the highest and second-highest values among all array elements. , Then the first one in the real-time graph i The principal directions of each feature point are calculated as follows:
[0010] in, Represents the main direction array The Middle i The element, represented as the element i The main direction of each feature point.
[0011] Furthermore, the radius scaling factor is adjusted twice, with each adjustment resulting in a larger value than the previous one.
[0012] Furthermore, the unique principal direction array of feature points in the real-time image and the baseline image. It is expressed as follows:
[0013] Among them, the three main direction arrays of the feature points in the real-time image are respectively , , The three main direction arrays of the feature points on the baseline map are respectively... , , ; 0.56, 0.77, and 1 are the three values of the radius coefficient, respectively.
[0014] Furthermore, based on the unique region main direction array of feature points... , No. i The absolute value of the principal direction difference for each pair of feature points is calculated as follows:
[0015] in, These represent the main direction arrays of the unique region. The first in i Each element value.
[0016] Furthermore, an index array is constructed and sorted using the absolute value of the difference in the main direction and preset matching effect indicators, including: For the first i Determine the absolute value of the principal direction difference between pairs of feature points and construct an index array. :
[0017] in m An adjustable threshold. For indicator array The first in i Each element value; First, remove the indicator array. The zero element value corresponds to the feature point pair. Then, the remaining feature point pairs are sorted according to their element values. The feature points are sorted from largest to smallest; finally, based on the required number of feature point pairs, a coarse matching is achieved by selecting from the sorted list from front to back.
[0018] A terminal device includes a processor, a memory, and a computer program stored in the memory; when the processor executes the computer program, it implements the heterogeneous scene matching point feature initial screening method based on the fused navigation information.
[0019] A computer-readable storage medium storing a computer program; when executed by a processor, the computer program implements a method for initial screening of heterogeneous scene matching point features based on fused navigation information.
[0020] Compared with the prior art, the present invention has the following technical features: 1. This invention provides a novel method and index for preliminary screening of coarse matching results of point features, which can complete the preliminary screening or sorting of coarse matching steps and improve the matching accuracy of the coarse matching process by 42%.
[0021] 2. Compared with the traditional method of determining based on feature point response values, the accuracy of the initial screening of matching point pairs will be higher, reducing the difficulty of the subsequent fine matching and shortening the image matching process time by 14% to 33%.
[0022] 3. The method proposed in this invention is applicable to all point features and has wide applicability. It can introduce a coarse matching step for all point features.
[0023] 4. This invention has better performance for heterogeneous image matching. The number of correct matching point pairs after coarse matching of heterogeneous images is generally low. When relying on traditional descriptor matching response values for screening, there will be a large number of false matches. The initial screening method proposed in this invention can improve the accuracy after screening. Attached Figure Description
[0024] Figure 1 This is a schematic flowchart of the method of the present invention; Figure 2 This is a set of input example images (coarse matching results of feature points between real-time image and reference image) in one embodiment of the method of the present invention; Figure 3 In response to Figure 2 The initial screening results obtained after processing using the method of the present invention. Detailed Implementation
[0025] This invention designs a method for initial screening of heterogeneous scene matching point features by fusing navigation information in the coarse matching process based on point features. This method can improve the accuracy of matching point pairs in the coarse matching step of heterogeneous scene matching, thereby significantly improving the accuracy of the subsequent fine matching step, while reducing the time consumption of the subsequent fine matching step. Due to the influence of factors such as scene, time, weather, and differences in information sources, the descriptor information of heterogeneous scene matching point features usually differs greatly, resulting in low accuracy of the coarse matching results. To address this problem, the algorithm designs an initial screening method for point features by fusing navigation information. First, the attitude and relative altitude information in the navigation information are used to establish the coordinate system between the real-time image and the reference image. Further, a weighted principal direction calculation method for matching point pairs is designed. Finally, considering that the difference between the correct principal directions of the two after fusing navigation information is 0, this invention designs a matching effect index for coarse matching point pairs based on the weighted principal direction.
[0026] Algorithm input: First, the attitude and relative height information in the navigation information are used to establish the coordinate system of the real-time image and the reference image. Then, feature point pairs are obtained by coarse matching of feature points between the real-time image and the reference image.
[0027] The method of this invention operates after the coarse matching step and before the fine matching step in the feature matching algorithm. In this scheme, the two matched images are referred to as the real-time image and the baseline image, respectively. In the feature point pair after coarse matching, the feature point set of the real-time image is... The set of feature points on the baseline image is ; For the set of feature points The first in i One feature point, For the set of feature points The first in i One feature point; N This represents the number of feature point pairs in the coarse matching. Based on this, the specific steps of the algorithm are as follows: Step 1: Real-time image and baseline image correction.
[0028] Based on the attitude and relative flight altitude calculated by the inertial navigation system of the image acquisition platform during real-time image acquisition, geographic information correction can be performed on the real-time image to ensure that the angle and scale of the corrected real-time image are consistent with the reference image; let the corrected real-time image be denoted as... . This step is a mature procedure in this field, as detailed in the article "A Geographic Information Correction Method Based on Inertial / Altimeter / Visual Fusion" published in the 10th issue of the Chinese Journal of Inertial Technology.
[0029] Step 2: Calculate the gradient magnitude map.
[0030] For the corrected real-time graph Calculate the first-order gradient magnitude map, denoted as... For the baseline diagram Calculate the first-order gradient magnitude plot . Step 3: Calculate the gradient direction pattern.
[0031] For the corrected real-time graph Calculate the first-order gradient pattern, denoted as... ; For the baseline diagram Calculate the first-order gradient pattern For two first-order gradient patterns and Perform pixel-by-pixel mapping and calculate the partition map respectively. The mapping process is as follows:
[0032] The calculation process can be described as follows: for the first-order gradient pattern Each value in the parameter is multiplied O The result is then divided by 360 and rounded; among which, O The threshold value is adjustable; in this embodiment, it is set to 24. rand This indicates the rounding process. Step 4: Calculate the weighted principal direction for each pair of feature points.
[0033] Regarding the first i Group feature point pairs , Its real-time graph after correction Compared with the baseline diagram The position coordinates on are denoted as follows: , Initialize the three main direction arrays of the feature points on the real-time image as follows: , , The three main direction arrays of the feature points on the initialization map are as follows: , , The array size is 1* N All array values are initialized to zero.
[0034] Set radius parameter =70, calculate the number on the real-time graph. i The steps for determining the principal direction of each feature point are as follows: Step 4.1, Initialize the histogram Theq element The histogram is 1* O An array of size , where all elements of the array are initialized to 0.
[0035] Step 4.2, set the radius coefficient =0.56, combined with the radius parameter Calculate the radius:
[0036] Step 4.3, for the corrected real-time graph The position coordinates of the upper feature point traverse distance smaller than radius R Position coordinates Statistical histogram:
[0037] in, Represents position coordinates In the partitioned graph corresponding to the real-time graph The position above, Indicates to Perform statistical histogram operations. Represents position coordinates The first-order gradient magnitude map corresponding to the real-time graph The position above, Histogram middle The element corresponding to the position.
[0038] Step 4.4, find the histogram The indices of the highest and second-highest values among all array elements. , Then the first one in the real-time graph i The principal directions of each feature point are calculated as follows:
[0039] in, Represents the main direction array The Middle i The element, represented as the element i The main direction of each feature point.
[0040] Step 4.5 Modify the radius coefficient in Step 4.2 =0.77, repeat steps 4.1 to 4.4 to calculate the... i New main direction for each feature point Similarly, let the radius proportionality coefficient... =1.00, calculate the first... iNew main direction for each feature point , Through the above steps, from coarse matching feature point pairs 1~ N Traversal i This allows us to obtain three main direction arrays for all feature points in the real-time image.
[0041] Step 4.6: Replace the real-time image with the reference image. Using the same method as steps 4.1-4.5, the three main direction arrays of all feature points on the reference image can be calculated. , , In step 4.1, the histogram corresponding to the real-time graph is initialized; in step 4.3, the partitioned graph corresponding to the baseline graph is used for calculation. and the first-order gradient magnitude map corresponding to the baseline map. .
[0042] Step 4.7: Weight the principal direction arrays of the three regions in the real-time image and the baseline image respectively, and calculate the unique principal direction arrays of feature points in the real-time image and the baseline image. :
[0043] Unlike traditional point feature principal direction calculation rules such as SURF, SIFT, and SAR-SIFT, this invention calculates multiple principal directions using circular regions of different radii centered on the feature point location, and further weights and fuses these multiple principal directions to construct a novel principal direction calculation rule.
[0044] Step 5: Calculate the difference between the matching points in the main direction.
[0045] Based on the unique region main direction array of feature points calculated in step 4 , No. i Absolute value of the principal direction difference between pairs of feature points:
[0046] in, These represent the main direction arrays of the unique region. The first in i Each element value.
[0047] Step 6: Calculate the matching effect index.
[0048] For the i Determine the absolute value of the principal direction difference between pairs of feature points and construct an index array. :
[0049] in mIn this embodiment, the threshold value is set to 60, which is an adjustable threshold. For indicator array The first in i Each element value; that is, this scheme uses auxiliary navigation information to correct the input image as prior information, and the specific description of this prior information is: the absolute value of the main direction difference of the feature point pair is 0.
[0050] Step 7: Coarse matching to complete the initial screening of feature point pairs.
[0051] Use the indicators from step 6 as the criteria for determining whether a match is successful; first, remove indicators from the array. The zero element value corresponds to the feature point pair. Then, the remaining feature point pairs are sorted according to their element values. The feature points are sorted from largest to smallest; finally, based on the required number of feature point pairs, a coarse matching is achieved by selecting from the sorted list from front to back.
[0052] Combination Figure 2 and Figure 3 The method of the present invention can perform preliminary screening of coarse matching results, which will greatly improve the accuracy of matching point pairs obtained in the coarse matching process, thereby reducing the complexity of the fine matching process in the back end; at the same time, the method proposed in the present invention has very low time consumption, which greatly improves the image matching time.
[0053] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. A method for initial screening of heterogeneous scene matching point features by fusing navigation information, characterized in that, include: Obtain the real-time image, and perform coarse matching of feature points between the real-time image and the baseline image to obtain the feature point sets of the real-time image and the baseline image. The real-time map is corrected for geographic information to align it with the baseline map; For the corrected real-time graph and the reference graph, calculate the corresponding first-order gradient magnitude graph and the corresponding first-order gradient direction graph. Pixel-by-pixel mapping is performed on the first-order gradient direction maps of the real-time map and the baseline map to obtain the corresponding partition maps; For real-time graphs, the radius is determined based on preset radius parameters and radius coefficients; For the position coordinates of feature points on the corrected real-time image, iterate through the position coordinates that are less than the radius from the feature point and construct a histogram, including: Initialize histogram The q element The histogram is 1* O An array of sizes; ； in, O This is an adjustable threshold; Represents position coordinates In the partitioned graph corresponding to the real-time graph The position above, Indicates to Perform statistical histogram operations. Represents position coordinates The first-order gradient magnitude map corresponding to the real-time graph The position above, Histogram middle The element corresponding to the position; The principal direction of each feature point is determined based on the histogram and recorded in the first region principal direction array; the radius scaling factor is adjusted twice to obtain the second and third region principal direction arrays; based on the three region principal direction arrays, the unique principal direction array of feature points in the real-time image is calculated; the same method is used to determine the unique principal direction array of the reference image. The absolute value of the difference between the principal directions is determined by using the unique principal direction arrays of the real-time graph and the baseline graph; An index array is constructed and sorted using the absolute value of the difference in the main direction and preset matching effect indicators, including: For the i Absolute value of the principal direction difference between pairs of feature points Make a judgment and construct an indicator array. : ； in m An adjustable threshold. For indicator array The first in i Each element value; First, remove the indicator array. The zero element value corresponds to the feature point pair. Then, the remaining feature point pairs are sorted according to their element values. The feature points are sorted from largest to smallest; finally, based on the required number of feature point pairs, they are selected from the sorted list from front to back to achieve coarse matching. Select the corresponding feature point pairs according to the required number of feature point pairs.
2. The method for initial screening of heterogeneous scene matching point features based on fused navigation information according to claim 1, characterized in that, Pixel-by-pixel mapping is performed on the first-order gradient direction maps of the real-time map and the baseline map to obtain the corresponding partition maps, including: For the corrected real-time graph Baseline diagram Calculate the first-order gradient pattern , ; For two first-order gradient patterns and Perform pixel-by-pixel mapping and calculate the partition map respectively. ; one of the first-order gradient patterns Mapping yields a partitioned map The process is as follows: For the first-order gradient pattern Each value in the parameter is multiplied O The result is then divided by 360 and rounded off.
3. The method for initial screening of heterogeneous scene matching point features based on fused navigation information according to claim 1, characterized in that, For real-time graphs, the radius is determined based on preset radius parameters and radius coefficients, including: Set radius parameter and radius coefficient Then the radius .
4. The method for initial screening of heterogeneous scene matching point features based on fused navigation information according to claim 1, characterized in that, The principal direction of each feature point is determined based on the histogram and recorded in the first region principal direction array, including: Find the histogram The indices of the highest and second-highest values among all array elements. , Then the first one in the real-time graph i The principal directions of each feature point are calculated as follows: ； in, Represents the main direction array Middle i The element, represented as the element i The main direction of each feature point.
5. The method for initial screening of heterogeneous scene matching point features based on fused navigation information according to claim 1, characterized in that, The radius scaling factor is adjusted twice, with each adjustment value being larger than the previous one.
6. The method for initial screening of heterogeneous scene matching point features based on fused navigation information according to claim 1, characterized in that, Unique principal direction array of feature points in real-time and baseline graphs It is expressed as follows: ； Among them, the three main direction arrays of the feature points in the real-time image are respectively , , The three main direction arrays of the feature points on the baseline map are respectively... , , ; 0.56, 0.77, and 1 are the three values of the radius coefficient, respectively.
7. A terminal device, comprising a processor, a memory, and a computer program stored in the memory; characterized in that, When the processor executes the computer program, it implements the heterogeneous scene matching point feature initial screening method according to any one of claims 1-6.
8. A computer-readable storage medium storing a computer program; characterized in that, When the computer program is executed by the processor, it implements the heterogeneous scene matching point feature initial screening method according to any one of claims 1-6.