Feature description and matching positioning method of ground penetrating radar image based on shape context

Abstract

The present invention provides a ground-penetrating radar image feature description and matching positioning method based on shape context, which belongs to the field of navigation positioning, including combining the two-dimensional B-scan image collected by the ground-penetrating radar along the road direction with the mileage and location information to spatially grid the image, so as to realize the alignment of the image with the mileage and location labels; designing a ground-penetrating radar image structure feature extraction method based on shape context to obtain the ground-penetrating radar image structure feature; using the ground-penetrating radar image structure feature extraction method based on shape context to construct a reference feature library of the area to be positioned; extracting the real-time ground-penetrating radar image structure feature; obtaining a coarse matching point pair; obtaining a location label in the reference feature library that matches the real-time structure feature, so as to realize the real-time positioning of the carrier. The present invention realizes vehicle positioning by matching the ground-penetrating radar image structure feature, and is suitable for the field of vehicle navigation positioning.

Description

Technical Field

[0001] The present invention belongs to the technical field of navigation and positioning, and in particular relates to a ground penetrating radar image feature description and matching positioning method based on shape context. Background Art

[0002] Ground Penetrating Radar (GPR) is a method of detecting underground structures by emitting and receiving electromagnetic waves. It is currently widely used in road flaw detection, engineering inspection and other fields. Since it can obtain images of shallow surface structures, it can be used as a new means for vehicles to perceive external environmental information. At the same time, since the underground environment remains basically unchanged on a time scale and is unique on a spatial scale, due to the detection principle of ground penetrating radar, it is less affected by bad weather such as heavy rain, dense fog, snow, sand and dust.

[0003] Most existing ground-penetrating radar positioning methods use a template-based matching strategy, that is, a reference map is built in advance, and the real-time collected images are evaluated for similarity with the reference map during positioning. If the similarity exceeds a predetermined threshold, the match is successful, and the location information corresponding to the reference library is output. Common similarity evaluation methods include normalized cross-correlation (NCC), mutual information (MI), sum-of-squared-differences (SSD), etc. Ground-penetrating radar is affected by ground reflection, antenna coupling, and device jitter, which makes images collected at different times at the same location easy to lose correlation. At the same time, the template matching-based method has the problem of being time-consuming. Recent papers have proposed methods for matching visual images, such as SIFT (Scale Invariant Feature Transform), ORB (Oriented FAST and Rotated BRIEF), etc., which are directly used for ground-penetrating radar image matching. Since ground-penetrating radar images lack clear corner points and boundary features compared to general optical images, the matching effect is not ideal. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a ground penetrating radar image feature description and matching positioning method based on shape context. By extracting the ground penetrating radar context structural features, on the one hand, the matching robustness is improved, on the other hand, the matching efficiency is effectively improved and the required storage space is reduced.

[0005] In order to achieve the above object, the present invention adopts the following technical scheme:

[0006] A ground penetrating radar image feature description and matching positioning method based on shape context includes the following steps:

[0007] Step 1: Combine the 2D B-scan image collected by the ground penetrating radar along the road direction with the mileage and location information to spatially grid the image and align the mileage and location labels;

[0008] Step 2: Design a structural feature extraction method based on shape context, use SURF to extract feature points of two-dimensional B-scan images, and build shape context feature descriptors around the feature points. The combination of the two constitutes the structural features of the ground penetrating radar image; SURF represents accelerated robust features;

[0009] Step 3: Use the context-based structural feature extraction method to build a benchmark feature library of the area to be located, which includes feature points, descriptors, mileage and location labels;

[0010] Step 4, obtaining a two-dimensional B-scan image of the track to be located in real time, and extracting real-time ground penetrating radar image structural features using a shape context-based structural feature extraction method;

[0011] Step 5: Use the nearest neighbor ratio to evaluate the similarity between the real-time structural features and the benchmark feature library to obtain a rough matching point pair;

[0012] Step 6: Eliminate mismatches in the rough matching point pairs through spatial geometric relationship verification and RANSAC to obtain position labels in the benchmark feature library that match the real-time structural features; RANSAC stands for random sampling consensus algorithm.

[0013] Furthermore, the step 1 includes: spatially gridding the two-dimensional B-scan image of the ground penetrating radar, specifically: combining the two-dimensional B-scan image of the time series with the mileage and position sequence which is also a time series, and realizing time alignment by combining linear interpolation. After alignment, each B-scan data has equal mileage intervals along the road direction, and mileage and position labels are obtained.

[0014] Furthermore, in step 2, the two-dimensional B-scan image is recorded as , whose length and width correspond to the length along the road direction and the length along the depth direction in the real world; in the computer, the two-dimensional B-scan image Is the size of The matrix of For the image Line Strength information of the column.

[0015] Furthermore, the method for extracting the context structure feature point set includes:

[0016] Step a: Extract two-dimensional B-scan image using SURF algorithm The feature point set in , , k ∈ ( 0 , K ] ,in, is the feature point set The feature points, with coordinates of ;

[0017] Step b: Set the threshold according to the average brightness of the image stripes ; For two-dimensional B-scan images Perform binarization to obtain a binary image :

[0018] ;

[0019] SURF feature point set Centered, binary image For statistical objects, a polar coordinate system is constructed; the polar coordinate system is divided into With fixed angle Divide the polar coordinate system evenly into Regions:

[0020] ;

[0021] in, and are the number of distance and angle divisions respectively; is the statistical radius;

[0022] Binarized image The non-zero points in are denoted as , , j ∈ ( 0 , J ] , J is the binary image The number of non-zero points in ; where for Middle non-zero points with coordinates ; Calculate feature points All non-zero points within the statistical radius R of the neighborhood Polar coordinates :

[0023] ;

[0024] in, and For feature points The horizontal and vertical coordinates of and For binary images The horizontal and vertical coordinates of the non-zero point in the middle;

[0025] Count the number of non-zero points falling into each region to form a descriptor , the matrix size is , normalize the descriptor;

[0026] Feature points combined with corresponding descriptors constitute the contextual structural features of the ground penetrating radar image.

[0027] Furthermore, in step 3, the reference feature library includes: feature point position coordinates, context structure feature descriptors corresponding to the feature points, and mileage information corresponding to the feature points; the feature point position coordinates include latitude, longitude and elevation.

[0028] Furthermore, step 4 includes using a ground penetrating radar to collect B-scan images in real time along the trajectory to be located, and using a ground penetrating radar image feature description method based on shape context to extract real-time ground penetrating radar image structural features.

[0029] Furthermore, in step 5, the feature database descriptor is D base = [ d base , 1 , d base , 2 , … , d base , T ] and real-time feature descriptors D current = [ d current , 1 , d current , 2 , … , d current , T ] to and to , to count the number of non-zero points falling into each region, the distance The calculation formula is:

[0030] ;

[0031] Calculate the distance between the real-time feature descriptor and the corresponding feature database descriptor, assuming that the closest and second closest distances are and ,when:

[0032] ;

[0033] That is, the ratio of the nearest distance to the second nearest distance is less than the threshold , then it is considered that a match is achieved between the two descriptors.

[0034] Furthermore, step 6 includes: randomly selecting three pairs of non-collinear matching feature point pairs respectively , and , forming two sets of triangles and , use the arccosine formula to calculate the interior angles of the two triangles; when the angle difference between the two triangles is less than a preset threshold, the two triangles are considered similar; then retain two sets of point pairs, select the fourth set of point pairs, and repeat the above steps; until four sets of point pairs are selected, check the spatial geometric relationship, and then perform RANSAC to eliminate false matches; finally, output the spatial position information corresponding to the successfully matched feature database as the positioning result.

[0035] The present invention is applied to the field of vehicle navigation and positioning, and has the following beneficial effects:

[0036] 1. The present invention can assist in supplementing conventional navigation and positioning means to meet the continuity and accuracy of positioning.

[0037] 2. The present invention adopts a feature-based matching method to extract contextual structural features in ground penetrating radar images for matching and positioning. Compared with the traditional template matching method, the matching efficiency is improved while ensuring the matching robustness.

[0038] 3. The present invention uses spatial geometric relationships to pre-check matching point pairs, reducing a large amount of calculations caused by random sampling of the RANSAC (random sampling consensus algorithm), and reducing the number of iterations and calculation time. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a schematic diagram of the ground penetrating radar installation;

[0040] Figure 2 It is a flow chart of a method for describing and matching ground penetrating radar image features based on shape context of the present invention;

[0041] Figure 3 It is a schematic diagram of context structure features;

[0042] Figure 4 This is a ground penetrating radar B-scan image. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0044] The present invention provides a method for describing and matching ground penetrating radar image features based on shape context. Figure 1The method shown installs the antenna of the ground penetrating radar to ensure that the driving track overlaps with the collection track. The shallow surface image collected by the ground penetrating radar is used as the original input of the algorithm of the present invention, and after calculation, the vehicle positioning result is finally output.

[0045] like Figure 2 As shown, a method for describing and matching ground penetrating radar image features based on shape context of the present invention comprises the following steps:

[0046] Step 1: The two-dimensional B-scan image collected by the ground penetrating radar along the road direction is spatially gridded in combination with the mileage and location information to achieve alignment of the two-dimensional B-scan image with the mileage and location labels;

[0047] Step 2: Design a structural feature extraction method based on shape context, use SURF (Speeded Up Robust Features) to extract B-scan image feature points, and build a shape context feature descriptor around the feature points. The combination of the two constitutes the structural features of the ground penetrating radar image.

[0048] Step 3: Use the context-based structural feature extraction method to build a benchmark feature library of the area to be located. The benchmark feature library includes feature points, descriptors, mileage and location labels.

[0049] Step 4: Acquire the two-dimensional B-scan image of the track to be located in real time, and extract the real-time ground penetrating radar image structural features using the shape context-based structural feature extraction method.

[0050] Step 5: Use the nearest neighbor ratio to evaluate the similarity between the real-time structural features and the benchmark feature library to obtain coarse matching point pairs.

[0051] Step 6: Eliminate mismatches in the rough matching point pairs through spatial geometric relationship verification and RANSAC (random sampling consensus algorithm), and finally obtain the position labels in the benchmark feature library that match the real-time structural features.

[0052] Specifically, step 1 includes: the ground penetrating radar is an air-coupled single-channel ground penetrating radar. The installation method is as follows Figure 1 As shown, the antenna of the ground penetrating radar is installed at the rear of the vehicle, with the antenna height of about 0.3m from the ground, and the center axis of the vehicle and the center axis of the antenna of the ground penetrating radar are ensured to be in the same straight line. The vehicle driving trajectory coincides with the ground penetrating radar detection trajectory. The antenna center frequency is 300MHz, the time window is set to 60ns, and the relative dielectric constant is 4.5. The detection depth is about 4m. The number of sampling points along the depth direction of each data is 512, and the value of each sampling point corresponds to the reflected signal intensity. The scanning interval of each data along the road direction is 0.016m. The mileage and position output frequency is 100Hz.

[0053] Specifically, step 3 includes: constructing a reference feature library including feature point coordinates (latitude, longitude, elevation), mileage, and feature descriptors. 2D B-scan image Is the size of The matrix of A two-dimensional B-scan image Middle Line Strength information of the column.

[0054] The context structure feature point set extraction method includes:

[0055] (a) Feature point extraction. 2D B-scan image extracted using SURF algorithm The feature point set in , , k ∈ ( 0 , K ] ,in is the feature point set The feature points, with coordinates of .

[0056] (b) Setting the threshold based on the average brightness of the image stripes . For 2D B-scan images Perform binarization to obtain a binary image :

[0057] ;

[0058] SURF feature point set Centered, binary image For statistical objects, construct a polar coordinate system, such as Figure 3 As shown. The polar coordinate system is fixed at a distance With fixed angle Divide the polar coordinate system evenly into Regions:

[0059] ;

[0060] in, and are the number of distance and angle divisions respectively. is the statistical radius. , and The value of jointly determines the descriptor's ability to describe the feature. Among them, the statistical radius The larger the value, the more information is used to describe the feature points, but the statistical radius Should be within an appropriate range, too small a statistical radius This will result in too little neighborhood information and weak descriptor feature description. This will cause the neighborhoods of adjacent feature points to overlap, resulting in adjacent feature points having similar descriptors and statistical radius. The value is 1m. and The value of determines the degree of detail of the feature description by the descriptor. and The larger the value, the more detailed the description of the feature, but too large and This will cause the features to become irrelevant. and The values ​​are 5 and 8;

[0061] Binarized image The non-zero points in , , j ∈ ( 0 , J ] , J is the binary image The number of non-zero points in ; for Middle non-zero points with coordinates Calculate feature points All non-zero points within the statistical radius R of the neighborhood Polar coordinates :

[0062] ;

[0063] in, and For feature points The horizontal and vertical coordinates of and For binary images The horizontal and vertical coordinates of the non-zero point in the.

[0064] Count the number of non-zero points falling into each region to form a descriptor , the matrix size is , and normalize the descriptor.

[0065] Specifically, step 3 includes: constructing a benchmark feature library of the area to be located using the context structure feature extraction method described in step 2. The benchmark feature library should include feature points and their descriptors as well as corresponding mileage and location tags. At the same time, the benchmark feature library should store features in segments in combination with road network information. Specifically, the benchmark feature library is segmented according to intersections to improve matching efficiency.

[0066] Specifically, step 4 includes: acquiring a two-dimensional B-scan image of the trajectory to be located in real time, and extracting real-time structural features using a structural feature extraction method based on shape context.

[0067] Specifically, step 5 includes: the size of the matching window and the search range determines the reliability of the positioning result to a certain extent. Too small a matching window and search range will cause the matching result to fall into the local optimum and mismatch, while too large a parameter setting will cause a large amount of redundant calculations. Therefore, the matching window is set to 20m and the search range is set to 100m.

[0068] The feature database descriptor is D base = [ d base , 1 , d base , 2 , … , d base , T ] ,in, to To count the number of non-zero points falling into each region, T is the number of regions counted; real-time feature descriptor D current = [ d current , 1 , d current , 2 , … , d current , T ] ,in, and To count the number of non-zero points falling into each region, T is the number of regions counted, then the distance The calculation formula is:

[0069] ;

[0070] Calculate the distance between the real-time feature descriptor and the corresponding feature database descriptor, assuming that the closest and second closest distances are and ,when:

[0071] ;

[0072] That is, the ratio of the nearest distance to the second nearest distance is less than the threshold When , it is considered that a match is achieved between the two descriptors. Determines the strictness of the match. Threshold The larger the value, the greater the chance of mismatch. Set to 0.6.

[0073] Specifically, step 6 includes: randomly selecting three pairs of non-collinear matching feature point pairs, namely , and , forming two sets of triangles and , use the arccosine formula to calculate the internal angles of the two triangles. According to the triangle similarity principle, if the internal angles of the two triangles correspond and are equal, the two triangles are similar. Considering that the ground penetrating radar image has slight deformation, when the angle difference between the two triangles is less than the preset threshold, the two triangles are considered similar. Then retain two sets of point pairs, select the fourth set of point pairs, and repeat the above steps. Until four sets of point pairs are selected, they are checked through spatial geometric relationships, and then RANSAC (random sampling consensus algorithm) is performed to eliminate mismatches. Because the ground penetrating radar image is affected by factors such as ground reflection and antenna coupling, the image will have nonlinear deformation. Even the relative positions of the correctly matched point pairs may change. Therefore, the similarity angle threshold should not be set too small. Finally, the spatial position information corresponding to the successfully matched feature database is output as the positioning result.

[0074] Figure 4 It is a two-dimensional B-scan image collected by the ground penetrating radar along the road direction, and the vertical direction is the depth direction of the ground downward.

[0075] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is 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 in the protection scope of the present invention.

Claims

1. A ground penetrating radar image feature description and matching positioning method based on shape context, characterized in that: The steps include: Step 1: Combine the 2D B-scan image collected by the ground penetrating radar along the road direction with the mileage and location information to spatially grid the image and align the mileage and location labels; Step 2: Design a structural feature extraction method based on shape context, use SURF to extract feature points of two-dimensional B-scan images, and construct a shape context feature descriptor around the feature points. The combination of the two constitutes the structural features of the ground penetrating radar image. SURF stands for Speeded Up Robust Features; Step 3: Use the context-based structural feature extraction method to build a benchmark feature library of the area to be located, which includes feature points, descriptors, mileage and location labels; Step 4, obtaining a two-dimensional B-scan image of the track to be located in real time, and extracting real-time ground penetrating radar image structural features using a shape context-based structural feature extraction method; Step 5: Use the nearest neighbor ratio to evaluate the similarity between the real-time structural features and the benchmark feature library to obtain a rough matching point pair; Step 6: Eliminate mismatches in the rough matching point pairs through spatial geometric relationship verification and RANSAC to obtain position labels in the benchmark feature library that match the real-time structural features; RANSAC stands for random sampling consensus algorithm.

2. The method for ground penetrating radar image feature description and matching positioning based on shape context according to claim 1, characterized in that: The step 1 comprises: spatially gridding the two-dimensional B-Scan image of the ground penetrating radar, specifically: combining the two-dimensional B-scan image of the time series with the mileage and position sequence of the time series by linear interpolation to achieve time alignment, after which each B-scan data is aligned with equal mileage intervals along the road direction, and mileage and position labels are obtained.

3. The method for describing and matching ground penetrating radar image features based on shape context according to claim 1, characterized in that: In step 2, the two-dimensional B-scan image is recorded as , whose length and width correspond to the length along the road direction and the length along the depth direction in the real world; in the computer, the two-dimensional B-scan image Is the size of The matrix of For the image Line Strength information of the column.

4. The method for describing and matching ground penetrating radar image features based on shape context according to claim 3, characterized in that: The extraction method of context structure feature point set includes: Step a: Extract two-dimensional B-scan image using SURF algorithm The feature point set in , , ,in, is the feature point set The feature points, with coordinates of ; Step b: Set the threshold according to the average brightness of the image stripes ; For two-dimensional B-scan images Perform binarization to obtain a binary image : ； SURF feature point set Centered, binary image For statistical objects, a polar coordinate system is constructed; the polar coordinate system is divided into With fixed angle Divide the polar coordinate system evenly into Regions: ； in, and are the number of distance and angle divisions respectively; is the statistical radius; Binarized image The non-zero points in are denoted as , , , J is the binary image The number of non-zero points in ; where for Middle non-zero points with coordinates ; Calculate feature points All non-zero points within the statistical radius R of the neighborhood Polar coordinates : ； in, and For feature points The horizontal and vertical coordinates of and For binary images The horizontal and vertical coordinates of the non-zero point in the middle; Count the number of non-zero points falling into each region to form a descriptor , the matrix size is , normalize the descriptor; Feature points combined with corresponding descriptors constitute the contextual structural features of the ground penetrating radar image.

5. The method for describing and matching ground penetrating radar image features based on shape context according to claim 1, characterized in that: In step 3, the reference feature library includes: feature point position coordinates, context structure feature descriptors corresponding to the feature points, and mileage information corresponding to the feature points; the feature point position coordinates include latitude, longitude and altitude.

6. The method for ground penetrating radar image feature description and matching positioning based on shape context according to claim 4, characterized in that: The step 4 includes using the ground penetrating radar to collect B-scan images in real time along the trajectory to be located, and using a ground penetrating radar image feature description method based on shape context to extract real-time ground penetrating radar image structural features.

7. The method for ground penetrating radar image feature description and matching positioning based on shape context according to claim 1, characterized in that: In step 5, the feature database descriptor is and real-time feature descriptors , to and to To count the number of non-zero points falling into each region, the distance The calculation formula is: ； Calculate the distance between the real-time feature descriptor and the corresponding feature database descriptor, assuming that the closest and second closest distances are and ,when: ； That is, the ratio of the nearest distance to the second nearest distance is less than the threshold , then it is considered that a match is achieved between the two descriptors.

8. The method for describing and matching ground penetrating radar image features based on shape context according to claim 1, characterized in that: The step 6 comprises: randomly selecting three pairs of non-collinear matching feature point pairs respectively , and , forming two sets of triangles and , use the arccosine formula to calculate the interior angles of the two triangles; when the angle difference between the two triangles is less than a preset threshold, the two triangles are considered similar; then retain two sets of point pairs, select the fourth set of point pairs, and repeat the above steps; until four sets of point pairs are selected, check the spatial geometric relationship, and then perform RANSAC to eliminate false matches; finally, output the spatial position information corresponding to the successfully matched feature database as the positioning result.

Images