A scene matching navigation method and system

Abstract

This invention discloses a novel scene matching navigation method, belonging to the field of visual navigation technology. First, an aerial map reference image is acquired, and the required training dataset is constructed through cropping and labeling. Then, the Yolov5 feature detection algorithm is used to detect features in the reference image, extracting semantic regions with clear features as semantic landmarks and constructing a semantic landmark library. Next, based on the nearest neighbor principle, a feature description triangle descriptor is constructed for each landmark and its surrounding landmarks. Geometric features are used to distinguish landmarks with the same semantic meaning, enhancing the specificity of landmark descriptions. Finally, a similarity measurement is performed between the triangle descriptors generated from real-time images and the triangle descriptor library to obtain the final relative position information and calculate navigation parameters. This invention, by constructing a new semantic feature descriptor, achieves effective image description, enhancing the accuracy, scene adaptability, and real-time performance of scene matching.

Description

Technical Field

[0001] This invention belongs to the field of visual navigation technology, and particularly relates to a scene matching navigation method and system. Background Technology

[0002] Image matching navigation, as an important auxiliary navigation method, boasts advantages such as all-weather operation, low cost, and high precision, and is widely used in navigation scenarios for aircraft and ground equipment. However, the process of matching and positioning based on a reference map and real-time images captured by the aircraft is time-consuming, making it difficult to deploy hardware systems and implement in practice. Therefore, improving the real-time performance of the matching process has become a current research hotspot.

[0003] Existing image matching techniques are mainly divided into two categories: feature-based matching methods and statistical information-based matching methods. Feature-based matching methods are currently the most widely used image matching algorithms. The basic principle is to describe certain features in an image using special descriptors, converting image information into descriptive descriptions. Then, the descriptors of the target image are matched with those of the reference image to obtain the final relative position information. Currently, widely used descriptors include point features, line features, and region features. After constructing descriptors, similarity measurements are used to determine image matching relationships, which has been well validated in some low-altitude, low-speed aircraft systems. However, for high-altitude aerial images with rich and complex features, the problem of excessively long real-time image descriptor construction time still exists, making it difficult to meet the real-time requirements of the system. Statistical information-based matching methods mainly focus on clustering and statistical analysis of image features such as grayscale, constructing feature vectors to describe image features, and then performing similarity measurements between images to obtain position information. For aerial image navigation scenarios, this method suffers from problems such as high computational cost and susceptibility to changes in scene lighting.

[0004] With the widespread application of deep learning algorithms in image processing, the real-time performance of many image processing tasks has been greatly improved, and a large number of deep learning-based matching algorithms have emerged in the field of image matching. For example, using neural networks for feature point extraction greatly reduces matching time, but it still suffers from low real-time performance when faced with a large number of pixels. Some algorithms also utilize the semantic features of images to quickly achieve coarse matching through object detection, and then use traditional algorithms to complete fine matching in the coarse matching region. However, these algorithms cannot withstand changes in image scale and lack the ability to distinguish between the same semantic targets. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a scene matching navigation method that achieves high accuracy, good scene adaptability, and short matching time during the matching process.

[0006] To achieve the above functions, the present invention adopts the following technical solution:

[0007] A scene matching navigation method includes the following steps:

[0008] S1. Obtain the map baseline image and create a semantic annotation dataset, and obtain real-time images;

[0009] S2. Train the Yolov5 network using the semantic annotation dataset to obtain a network structure that can detect special semantic regions of the target, detect special semantic regions of the map reference image, and obtain the landmark description of the map reference image.

[0010] S3. Delete false checkpoints and build a road sign database;

[0011] S4. Construct a feature triangle descriptor for each landmark and its surrounding landmarks based on the nearest neighbor principle;

[0012] S5. A triangle descriptor sub-library is constructed by optimizing redundant terms and selecting the optimal angle.

[0013] S6. Based on the triangle descriptors generated from the real-time images and the triangle descriptor library, a similarity measurement is performed. The optimal error standard is used to obtain the matching triangle descriptors in the library. The relative coordinate information of the real-time region is converted into the absolute coordinate information of the triangle descriptor library entries, and the current absolute position information of the aircraft is calculated.

[0014] As a preferred option, step S1 specifically involves first cropping the map reference image according to a certain pixel size; then performing additional image cropping at the seams of the cropped area to ensure that each semantic region can be completely cropped; finally, selecting the playground and crossroads areas in the urban scene as special semantic regions, and using Labelme software to annotate the cropped images to obtain a semantic annotation dataset.

[0015] Preferably, in step S2, the descriptive terms for the landmark points in the map reference image are represented as follows:

[0016]

[0017] in, Indicates the current landmark number; Indicates the first One road sign; Indicates the category of the current landmark; This represents the confidence score of the current landmark. The confidence score characterizes the similarity score between the current detection area and the area in the label data. The higher the score, the higher the similarity. , This indicates the absolute x-coordinate and absolute y-coordinate of the current landmark on the base map.

[0018] Preferably, in step S3, deleting false detection points involves: identifying and deleting road sign entries whose confidence scores for map reference image road sign descriptions are lower than a set value; constructing the road sign database involves: storing qualified road sign entries as .npy format files to construct the road sign database.

[0019] Preferably, step S4 involves the following steps:

[0020] Sort the landmarks according to their coordinate values, and select the landmarks to be described as reference points in order; calculate the Euclidean distance between the reference points and the description points, and select the description points based on the distance values;

[0021] Using the reference point as the vertex of the triangle and the description point as the other two vertices of the triangle, a feature triangle is constructed, and semantic information terms, triangle shape information, and triangle position of the feature triangle are constructed to build a triangle descriptor.

[0022] Preferably, in step S4, selecting the descriptive point specifically means: selecting a distance less than a threshold distance. The point is used as the description point. The value is the diagonal length of the rectangular region of the real-time detection image.

[0023] Preferably, in step S4, the semantic information entries of the feature triangle are:

[0024] The semantic information obtained from the three endpoints of the feature triangle through the semantic detection network is statistically analyzed, and then sorted according to the size relationship of the interior angles formed by the three endpoints as vertices, as shown below:

[0025]

[0026] In the formula, the subscript Indicates the first A feature triangle term, The semantic category of the largest corner vertex, Indicates the semantic category of the second largest vertex. The semantic category of the smallest corner vertex;

[0027] The triangle shape information:

[0028] The shape of the triangle is described by selecting two interior angles of the characteristic triangle, represented as:

[0029]

[0030] In the formula, Indicates the first The largest interior angle of the characteristic triangle, Indicates the first The second largest interior angle of a characteristic triangle;

[0031] The position of the triangle:

[0032] The position of a triangle is described by its circumcenter, represented as:

[0033]

[0034] In the formula, Indicates the first The absolute x-coordinate of the circumcenter of each characteristic triangle Indicates the first The absolute ordinate of the circumcenter of each characteristic triangle;

[0035] The triangle descriptor is represented as follows:

[0036] .

[0037] Preferably, in step S5, the redundant entries are optimized by deleting duplicate triangle descriptors that describe the same feature triangle; the optimal angle selection is to delete cases where the interior angles of the triangle are too large or too small, and only retain triangle descriptor entries whose interior angles are within a preset range; and the qualified triangle descriptor entries are stored as .npy format files to construct a triangle descriptor database.

[0038] Preferably, in step S6, the similarity measure is:

[0039] Different values ​​are assigned to different categories, and the feature triangle descriptors in the database are evaluated sequentially. The matching is successful if the category label={label1,label2,label3} in the image matches the category of the triangle descriptor generated from the actual image; otherwise, the matching fails.

[0040] Preferably, the optimal error criterion is as follows: when a triangle descriptor matches successfully, the interior angles are judged to be equal based on the angle error. If the angle error is less than a set threshold, the angles are considered equal; otherwise, they are considered unequal, and the matching fails. The angle error is expressed as follows:

[0041]

[0042] in, , This represents the largest and second largest interior angles of triangle descriptors in the triangle descriptor library. , Indicates real-time Figure 3The largest internal angle and the second largest internal angle of the angular descriptor;

[0043] Among the terms that successfully match, the triangular descriptor from the library with the smallest matching error is selected as the final result, represented as:

[0044]

[0045] in, This represents the matching error of all successfully matched triangle descriptors. This represents the minimum value among all errors.

[0046] The present invention also provides a scene matching navigation system, comprising:

[0047] The landmark acquisition module uses a self-made semantic annotation dataset to train a Yolov5 network to obtain a network structure that can detect special semantic regions. It then performs special semantic region detection on the map reference image to obtain landmark descriptions of the map reference image.

[0048] The landmark database construction module identifies landmark entries with a confidence score of less than 0.5 in the map reference image as false points and deletes them, thus constructing the landmark database.

[0049] The feature triangle descriptor construction module constructs a feature triangle descriptor for each landmark point and its surrounding landmark points based on the nearest neighbor principle.

[0050] The triangle descriptor library construction module uses redundant entries optimization and optimal angle selection to construct the triangle descriptor library.

[0051] The triangle descriptor matching module performs similarity measurement between triangle descriptors generated from real-time images and a triangle descriptor library. It then uses the optimal error standard to obtain matching triangle descriptors from the library and finally obtains the relative position information to calculate navigation parameters.

[0052] The present invention adopts the above technical solution, and its significant technical effects compared with the prior art are as follows:

[0053] (1) In the acquisition of road signs, the present invention uses the Yolov5 algorithm to extract semantic information of image road signs and describe the image, making full use of the semantic information in the image, realizing the dimensionality reduction description of image features, and enhancing the scene adaptability of image features;

[0054] (2) This invention constructs triangle descriptors by the nearest principle, enriches the description of landmark points, and constructs a triangle descriptor library, thereby enhancing the globality, scale invariance, and specificity of image semantic description for the same semantic target; and in the process of constructing triangle descriptors, the redundant term optimization method is adopted to reduce the redundancy of the triangle descriptor library, and the robustness of triangle descriptors is improved by the angle optimization principle.

[0055] (3) The present invention adopts the optimal error matching principle in the similarity measurement process, which improves the accuracy of scene matching;

[0056] (4) By introducing the triangle descriptor, the present invention constructs a new semantic feature descriptor, which has the advantages of high accuracy, good scene adaptability and short matching time in the scene matching process. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the overall steps of one embodiment of the present invention;

[0058] Figure 2 This is a flowchart of the detection network training process according to an embodiment of the present invention;

[0059] Figure 3 This is a flowchart illustrating the construction process of a semantic landmark library according to an embodiment of the present invention;

[0060] Figure 4 This is a flowchart of the feature triangle descriptor construction process according to an embodiment of the present invention;

[0061] Figure 5 This is a schematic diagram showing the positional relationship between a reference point and a description point according to an embodiment of the present invention;

[0062] Figure 6 This is a schematic diagram of the geometric configuration of a feature triangle descriptor according to an embodiment of the present invention;

[0063] Figure 7 This is a flowchart illustrating the construction process of a triangle descriptor library according to an embodiment of the present invention;

[0064] Figure 8 This is a map library reference image according to an embodiment of the present invention;

[0065] Figure 9 (a) is the original image of an embodiment of the present invention;

[0066] Figure 9 (b) is a semantic tag image according to an embodiment of the present invention;

[0067] Figure 10 (a) is the semantic detection result of an embodiment of the present invention;

[0068] Figure 10 (b) is a landmark description of an image according to an embodiment of the present invention;

[0069] Figure 11 This is a reference image feature point library according to an embodiment of the present invention;

[0070] Figure 12 (a1)-(a4) are real-time images of one embodiment of the present invention;

[0071] Figure 12 (b1)-(b4) are real-time image detection results of one embodiment of the present invention;

[0072] Figure 12 (c1)-(c4) are the matching results of the scene matching navigation method in one embodiment of the present invention. Detailed Implementation

[0073] The specific implementation of the present invention will be further described below with reference to the accompanying drawings:

[0074] This invention achieves dimensionality reduction in image feature description by introducing a triangular descriptor, enhancing the scene adaptability of image features. Furthermore, it enriches the description of landmark points, enhancing their globality, scale invariance, and specificity for the same semantic target. Additionally, it reduces the redundancy of the triangular descriptor library through redundant term optimization and improves the robustness of the triangular descriptors through angle optimization principles. The optimal error matching principle is employed in the similarity measurement process, improving the accuracy of scene matching. Ultimately, it achieves high accuracy, good scene adaptability, and short matching time in scene matching.

[0075] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0076] The present invention mentions a scene matching navigation method, such as... Figure 1 As shown, it includes the following steps:

[0077] S1. Obtain a reference map from the airborne map database. In this embodiment, the reference image used in the map database is as follows: Figure 8 As shown.

[0078] S2. Cropping and labeling of special semantic regions using the map reference image to obtain a semantically labeled dataset, and training the Yolov5 detection network as follows: Figure 2 As shown, the specific steps are as follows:

[0079] S201. Crop the map reference image sequentially according to a pixel size of 512×512;

[0080] S202. Perform additional image cropping at the connection points of the cropping regions in S201 to ensure that each semantic region can be completely cropped and to guarantee the integrity of the target.

[0081] S203. Select the playground and crossroads areas in the urban scene as special semantic regions. Annotate the cropped images using Labelme software to obtain a dataset with annotated special semantic regions. In this embodiment, the original image and the semantically labeled image are as follows: Figure 9 As shown in (a) and (b);

[0082] S204. Input the dataset with special region annotations into the Yolov5 detection network for network training to obtain a network structure that can detect special semantic regions of the target.

[0083] S3. Detect special semantic regions in the map reference image, obtain landmark descriptions from the map reference image, and construct a semantic landmark library. (The process is as follows...) Figure 3 As shown, the specific steps are as follows:

[0084] S301. After cropping the map reference image, input the resulting dataset into the trained Yolov5 network for special semantic region detection to obtain landmark descriptions of the map reference image. Then, perform term descriptions on the landmarks. The specific formulas for the landmark description terms are shown below:

[0085]

[0086] in, Indicates the current landmark number. Indicates the first One road sign, Indicates the category of the current landmark. This represents the confidence score of the current landmark. , This represents the absolute x-coordinate and absolute y-coordinate of the current landmark in the base map; a total of N landmarks are obtained. In this embodiment, the semantic detection results and the landmark descriptions in the image are as follows: Figure 10 As shown in (a) and (b).

[0087] S302. Determine that road sign entries in the road sign database with a confidence score lower than 0.5 as false detection points and delete them, reducing the number of road signs from N to M.

[0088] S303. Store the qualified landmark entries as .npy format files to construct a landmark database. In this embodiment, the landmark database is described on the image as follows: Figure 11 As shown.

[0089] S4. Construct a feature triangle descriptor for each landmark and its surrounding landmarks based on the nearest neighbor principle. The process is as follows: Figure 4 As shown, the specific steps are as follows:

[0090] S401. Sort the road signs according to their x-coordinate values ​​to obtain an ordered set of road signs;

[0091] S402. First, select the landmark points to be described in sequence, which will be referred to as reference points. Then, evenly distribute 50 points before and after each reference point and calculate the Euclidean distance between these points and the description points. Sort the distances and select those with a distance less than a threshold. The point is used as the description point. The value is the diagonal length of the rectangular region of the real-time detection image; the positional relationship between the reference point and the description point is as follows: Figure 5 As shown:

[0092] Triangular points represent selected reference points, rectangular points represent description points that meet the selection criteria, and circular points represent landmark points that do not meet the criteria. The radius of the circular area is the threshold distance. . The value is the diagonal length of the rectangular region of the real-time detection image.

[0093] S403. Construct a feature triangle by combining the reference point and the selected descriptive points to describe the image information of the current region; wherein, the reference point serves as the vertex of the feature triangle, and the remaining descriptive points are arranged and combined to serve as endpoints 1 and 2 of the feature triangle, respectively, to form the feature triangle; the configuration of the feature triangle is as follows: Figure 6 As shown:

[0094] When there are 4 descriptive points, a total of 6 can be formed. Triangles, respectively ;

[0095] S404. The interior angles formed by the three endpoints of the characteristic triangle are counted and sorted. The relationship between the three interior angles is as follows:

[0096]

[0097] Among them, satisfying ;

[0098] S405. The semantic information obtained from the three endpoints of the feature triangle through the semantic detection network is statistically analyzed, and sorted according to the size relationship of the interior angles formed by the three endpoints as vertices to form semantic information entries for the feature triangle, as shown below:

[0099]

[0100] in, Indicates the index of the current feature triangle. Indicates the first A feature triangle term, The semantic category of the largest corner vertex, Indicates the semantic category of the second largest vertex. The semantic category of the smallest corner vertex;

[0101] S406. Select two interior angles of the feature triangle to describe the triangle shape. The specific triangle shape information is shown below:

[0102]

[0103] in, Indicates the first The largest interior angle of the characteristic triangle, Indicates the first The second largest interior angle of a characteristic triangle;

[0104] S407. Calculate the circumcenter of the triangle. The specific calculation process is shown below:

[0105]

[0106]

[0107]

[0108]

[0109] in, These represent the absolute coordinates of the three endpoints. Indicates the first The absolute x-coordinate of the circumcenter of each characteristic triangle Indicates the first The absolute ordinate of the circumcenter of each characteristic triangle;

[0110] S408. Select the circumcenter of the triangle to describe its position, as shown below:

[0111]

[0112] in, Indicates the first The absolute x-coordinate of the circumcenter of each characteristic triangle Indicates the first The absolute ordinate of the circumcenter of each characteristic triangle;

[0113] S409. Integrate the information from S405, S406, and S408 to construct a triangular descriptor. The triangular descriptor information is converted into the specific representation of the terms as shown below:

[0114]

[0115] S410. Repeat S402~S409 for all landmarks to obtain all triangle descriptor entries.

[0116] S5. Construct a triangle descriptor sub-library using redundant term optimization and optimal angle selection methods. The flowchart is as follows: Figure 7 As shown, the specific steps are as follows:

[0117] S501. After completing S4, a total of P triangle descriptors are obtained. Redundancy optimization is used to delete duplicate triangle descriptors representing a feature triangle.

[0118] S502. By optimizing the angles, cases where the interior angles of the triangles are too large or too small are deleted, and only triangle descriptor entries with interior angles between 20° and 160° are retained; a total of Q triangle descriptors are obtained;

[0119] S503. Valid triangle descriptor entries are stored as .npy format files and used to construct a triangle descriptor database.

[0120] S6. Similarity measurement is performed between the triangle descriptors generated from the real-time image and the triangle descriptor library. The optimal error criterion is used to obtain the matching triangle descriptors from the library, ultimately yielding relative position information. Navigation parameters are then calculated. The process is as follows: Figure 1 As shown, the specific steps are as follows;

[0121] S601. Real-time images captured by the aircraft are used to detect landmarks through a trained Yolov5 semantic detection network to obtain a real-time landmark set.

[0122] S602. Determine the number of landmarks in the real-time image. When the number of landmarks is greater than or equal to 3, construct a feature triangle descriptor. When the number of landmarks is less than 3, the requirement for constructing a feature triangle cannot be met, and the matching algorithm in this paper cannot be used.

[0123] S603. When the number of road signs is greater than or equal to 3, perform a similarity measurement on the triangle descriptors generated from the real-time image and the triangle descriptor library.

[0124] S604. Assign different values ​​to different categories (assign label=0 when the category is playground, and label=1 when the category is intersection), and sequentially judge the feature triangle descriptors in the database. The system checks if the categories label={label1,label2,label3} in the image match the categories of the triangle descriptors generated from the actual image. If they match, the process proceeds to the next step; otherwise, the matching fails. If they don't match, the match fails; if they do match, it's considered a yes, and further checks are performed. Are they consistent? If so... If they don't match, the match fails; if they do match, it's considered a yes, and further checks are performed. Are they consistent? If so... If they do not match, the match fails; If the conditions are met, the system is deemed to be correct and the process jumps to S605.

[0125] S605. Determine if the interior angles are equal based on the angle error. If the error is less than a set threshold, they are considered equal; otherwise, they are considered unequal, and the matching fails. The specific angle error is as follows:

[0126]

[0127] in, , This represents the largest and second largest interior angles of triangle descriptors in the triangle descriptor library. , Indicates real-time Figure 3 The largest interior angle and the second largest interior angle of the angular descriptor, The threshold is set to 1°;

[0128] S606. Among the successfully matched terms, select the triangular descriptor from the library with the smallest matching error as the final result. The error selection method is as follows:

[0129]

[0130] in, This represents the matching error of all successfully matched triangle descriptors. This represents the minimum value among all errors;

[0131] S607. The successfully matched term is the first one in the currently successfully matched triangle descriptor sub-term. A triangular descriptor;

[0132] S608. Convert the relative coordinate information of the actual shooting area into the absolute coordinate information of the triangle descriptor library entries. Assuming the image center is the aircraft's position, calculate the aircraft's current absolute position information. The matching results in multi-lighting scenes are as follows: Figure 12 As shown, where Figure 12 (a1)-(a4) are real-time images. Figure 12 (b1)-(b4) represent the detection results corresponding to the real-time images. Figure 12 (c1)-(c4) show the matching results of this scene matching navigation method. As can be seen from the figures, the scene matching method disclosed in this application has high matching accuracy and strong scene adaptability, providing a new approach to image description and matching.

[0133] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A scene matching navigation method, characterized in that, include: S1. Obtain the map baseline image and create a semantic annotation dataset, and obtain real-time images; S2. Train the Yolov5 network using the semantic annotation dataset to obtain a network structure that can detect special semantic regions of the target, detect special semantic regions of the map reference image, and obtain the landmark description of the map reference image. S3. Delete false checkpoints and build a road sign database; S4. Construct a feature triangle descriptor for each landmark and its surrounding landmarks based on the nearest neighbor principle; S5. A triangle descriptor library is constructed using redundant term optimization and optimal angle selection. Redundant term optimization involves deleting duplicate triangle descriptors describing the same feature triangle. Optimal angle selection involves deleting triangle descriptors with excessively large or small interior angles, retaining only those with interior angles within a preset range. The qualified triangle descriptor terms are stored as .npy format files to construct a triangle descriptor database. S6. Based on the triangle descriptors generated from the real-time images and the triangle descriptor library, a similarity measurement is performed. The optimal error standard is used to obtain the matching triangle descriptors in the library. The relative coordinate information of the real-time region is converted into the absolute coordinate information of the triangle descriptor library entries, and the current absolute position information of the aircraft is calculated.

2. The scene matching navigation method according to claim 1, characterized in that, Specifically, step S1 involves first cropping the map reference image to a certain pixel size; then performing additional image cropping at the seams of the cropped area to ensure that each semantic region can be completely cropped; finally, selecting the playground and crossroads areas in the urban scene as special semantic regions, and using Labelme software to annotate the cropped images to obtain a semantic annotation dataset.

3. The scene matching navigation method according to claim 1, characterized in that, In step S2, the descriptive terms for the landmark points in the map reference image are represented as follows: ， in, Indicates the current landmark number; Indicates the first One road sign; Indicates the category of the current landmark; This represents the confidence score of the current landmark. The confidence score characterizes the similarity score between the current detection area and the area in the label data. The higher the score, the higher the similarity. , This indicates the absolute x-coordinate and absolute y-coordinate of the current landmark on the base map.

4. The scene matching navigation method according to claim 1, characterized in that, In step S3, deleting false detection points involves: identifying and deleting road sign entries whose confidence scores for map reference image road sign descriptions are lower than a set value; constructing the road sign database involves: storing qualified road sign entries as .npy format files to construct the road sign database.

5. The scene matching navigation method according to claim 1, characterized in that, In step S4, the specific steps are as follows: Sort the landmarks according to their coordinate values, and select the landmarks to be described as reference points in order; calculate the Euclidean distance between the reference points and the description points, and select the description points based on the distance values; Using the reference point as the vertex of the triangle and the description point as the other two vertices of the triangle, a feature triangle is constructed, and semantic information terms, triangle shape information, and triangle position of the feature triangle are constructed to build a triangle descriptor.

6. The scene matching navigation method according to claim 5, characterized in that, In step S4, selecting the descriptive point specifically involves selecting a distance less than a threshold distance. The point is used as the description point. The value is the diagonal length of the rectangular region of the real-time detection image.

7. The scene matching navigation method according to claim 5, characterized in that, In step S4, the semantic information entries of the feature triangle are: The semantic information obtained from the three endpoints of the feature triangle through the semantic detection network is statistically analyzed, and then sorted according to the size relationship of the interior angles formed by the three endpoints as vertices, as shown below: ， In the formula, the subscript Indicates the first A feature triangle term, The semantic category of the largest corner vertex, Indicates the semantic category of the second largest vertex. The semantic category of the smallest corner vertex; The triangle shape information: The shape of the triangle is described by selecting two interior angles of the characteristic triangle, represented as: ， In the formula, Indicates the first The largest interior angle of the characteristic triangle, Indicates the first The second largest interior angle of a characteristic triangle; The position of the triangle: The position of a triangle is described by its circumcenter, represented as: ， In the formula, Indicates the first The absolute x-coordinate of the circumcenter of each characteristic triangle Indicates the first The absolute ordinate of the circumcenter of each characteristic triangle; The triangular descriptor is represented as follows: 。 8. The scene matching navigation method according to any one of claims 1-7, characterized in that, In step S6, the similarity metric is: Different values ​​are assigned to different categories, and the feature triangle descriptors in the database are evaluated sequentially. The matching is successful if the category label={label1,label2,label3} in the image matches the category of the triangle descriptor generated from the actual image; otherwise, the matching fails. The optimal error criterion is as follows: when a triangle descriptor matches successfully, the interior angles are judged to be equal based on the angle error. If the angle error is less than a set threshold, the angles are considered equal; otherwise, they are considered unequal, and the match fails. The angle error is expressed as follows: ， in, , This represents the largest and second largest interior angles of triangle descriptors in the triangle descriptor library. , This represents the largest and second largest interior angles of the triangle descriptor in the real-time graph; Among the terms that successfully match, the triangular descriptor from the library with the smallest matching error is selected as the final result, represented as: ， in, This represents the matching error of all successfully matched triangle descriptors. This represents the minimum value among all errors.

9. A scene matching navigation system, characterized in that, include: The landmark acquisition module uses a self-made semantic annotation dataset to train a Yolov5 network to obtain a network structure that can detect special semantic regions. It then performs special semantic region detection on the map reference image to obtain landmark descriptions of the map reference image. The landmark database construction module identifies landmark entries in the map reference image whose confidence scores are lower than a preset value as false points and deletes them, thereby constructing the landmark database. The feature triangle descriptor construction module constructs a feature triangle descriptor for each landmark point and its surrounding landmark points based on the nearest neighbor principle. The triangle descriptor library construction module employs redundant term optimization and optimal angle selection to build the triangle descriptor library. Redundant term optimization involves deleting duplicate triangle descriptors describing the same feature triangle. Optimal angle selection involves deleting triangle descriptors with excessively large or small interior angles, retaining only those with interior angles within a preset range. The qualified triangle descriptor entries are stored as .npy format files, forming a triangle descriptor database. The triangle descriptor matching module measures the similarity between triangle descriptors generated from real-time images and a triangle descriptor library. It then uses the optimal error criterion to obtain matching triangle descriptors from the library and finally obtains the relative position information to calculate navigation parameters.

Images