SAR image feature matching method based on inertial navigation error correction

Abstract

The present invention relates to a SAR image feature matching method based on inertial navigation error information correction. The method uses matching points obtained according to a search strategy to estimate the approximate position of a real-time image in a reference image. The inertial navigation error information is then used to design a threshold for the difference between the relative position of the matching coordinates and the estimated position. Finally, the threshold is used to screen the difference between the position of the matching points and the estimated position, thereby removing a large number of abnormal mismatched points. The method is used to remove a large number of mismatches in the matching points obtained after performing nearest neighbor search and matching on feature points, thereby improving the robustness of the SAR image feature matching algorithm. The method solves the problem that a large number of mismatched points exist in the matching points before the affine transformation matrix is ​​calculated using a random sampling consistency algorithm, which causes the calculation of the affine transformation matrix to be time-consuming or even erroneous.

Description

Technical Field

[0001] The present invention belongs to the field of image technology and relates to a new SAR image feature matching algorithm based on inertial navigation error correction. The algorithm is used to remove mismatched point pairs in matching points obtained by nearest neighbor search based on feature vectors calculated according to phase consistency, so as to improve the accuracy of feature matching and reduce matching time. Background Art

[0002] Because it is unaffected by weather, lighting, and even surface obstructions, synthetic aperture radar (SAR) has been widely used on platforms such as satellites, drones, and missiles. SAR can also be used for imaging, and image matching is one application of SAR images. Image matching refers to the process of spatially aligning two images acquired by different sensors, at different times, or from different perspectives. In recent years, the development of imaging technologies such as visible light, infrared, microwave, and laser has generated a large amount of ground object data images. Aligning images acquired by different sensors or at different times can yield complementary information, and is therefore widely used in areas such as surface vegetation analysis, change detection, image fusion, and image retrieval. Image matching can also be used to assist in the correction of navigation and positioning systems.

[0003] Common navigation and positioning systems include the Inertial Navigation System (INS) and the Satellite Navigation System (GPS). INS utilizes its own inertial components to integrate position, independent of external information, and therefore possesses excellent anti-interference capabilities. However, positioning errors in INS accumulate over time, and prolonged flight can severely impact the positioning and velocity accuracy of INS. Satellite navigation systems utilize the time differences between signals received by the platform from multiple satellites or base stations for differential positioning, offering real-time performance and high positioning accuracy. However, satellite navigation systems rely on external electromagnetic information and are susceptible to interference and disruption during wartime, necessitating the development of navigation systems designed for satellite denial. Existing navigation systems typically combine INS with other navigation systems to create integrated navigation systems to improve navigation and positioning accuracy.

[0004] Combining SAR image matching with an inertial navigation system and using image matching to assist in correcting the inertial navigation system is a new integrated navigation system. This navigation system leverages the all-day, all-weather capabilities of the SAR system. While the SAR platform is in flight, it images the ground area and aligns the resulting image with a reference image stored on the platform to obtain matching relationships for key points. Since the position and elevation relationships of each point on the reference image are known, the latitude, longitude, and elevation information of each feature point in the SAR image can be obtained. Combined with the platform's heading, elevation, and radar beam parameters, the platform's position and attitude can be reverse-derived. This position and attitude information can be used to correct for the accumulated errors in the inertial navigation system caused by long-term flight, thereby improving the platform's positioning and velocity measurement accuracy.

[0005] Image matching algorithms are mainly divided into two categories: grayscale-based image matching algorithms and feature-based image matching algorithms. Grayscale-based image matching algorithms use the grayscale information between the images to be matched to perform matching, while feature-based image matching algorithms extract feature information from the images to be matched and match these features.

[0006] 1. Grayscale-based image matching algorithm:

[0007] Grayscale-based image matching algorithms such as Figure 1 As shown, a template window is slid across the reference image. Within each template window, the similarity between the live image (obtained by the SAR imaging system on the platform) and the reference image (pre-stored) is calculated based on the grayscale information. The location with the highest similarity is the matching location of the live image in the reference image. Grayscale information can be in the spatial or frequency domain. Common similarity measurement strategies include absolute square difference, cross-correlation, normalized cross-correlation, and mutual information. The absolute square difference method calculates the grayscale difference between the live image and the template window, defining the location with the minimum grayscale difference as the matching location. Therefore, it is sensitive to image grayscale distortion. The cross-correlation and normalized cross-correlation methods calculate the correlation coefficient or normalized cross-correlation coefficient between the live image and the template window, defining the location with the maximum correlation coefficient as the matching location. Therefore, they are resistant to linear grayscale distortion between the images. The mutual information method calculates the mutual information between the live image and the template window, defining the location with the maximum mutual entropy as the matching location. However, due to its high computational complexity and low computational efficiency, it is rarely used in practical engineering.

[0008] In general, grayscale-based image matching algorithms directly utilize image grayscale information. While computationally simple and straightforward, these algorithms are not very robust. When nonlinear grayscale distortion or geometric deformation exists between images, the matching performance of the algorithms drops sharply. Furthermore, grayscale-based image matching algorithms can only provide a single matching position. When the matching position is inaccurate or incorrect, it can severely impact the performance of subsequent positioning algorithms.

[0009] 2. Feature-based image matching algorithm:

[0010] Feature-based image matching algorithms such as Figure 2 As shown in Figure 1, this type of algorithm typically involves three steps: feature point detection, feature point description, and feature point matching. In theory, feature detection should filter out key points with distinct characteristics, such as edges, corners, and line intersections, from the image pair to be registered. Feature description characterizes these key points based on their relationships with surrounding pixels. The feature descriptions of these key points should be similar. Feature matching involves matching the feature point set of the baseline image with the feature point set of the live image based on a specific similarity metric.

[0011] 1. Feature point detection

[0012] This algorithm uses the Radiation-Variation Insensitive Feature Transform (RIFT) algorithm to extract feature points. RIFT first forms a maximum index map based on phase consistency. The formation process is as follows:

[0013] The input image is convolved with the log-Gabor filter, which is defined in the frequency domain as (1).

[0014]

[0015] Where ω0 is the center frequency, σ ω is the bandwidth parameter about the center frequency. The log-Gabor filter does not have a time domain analytical solution, but it can be obtained by performing an inverse Fourier transform (IFFT) on equation (1). Assume that the odd and even parts of the IFFT of the log-Gabor filter are h when the scale is s and the direction is o. e so (x,y) and h o so (x,y), then the convolution result of the input image is

[0016] [E so (x,y),O so (x,y)]=[I(x,y)*h eso (x,y),I(x,y)*h o so (x,y)] (2)

[0017] Where I(x,y) represents the input image. Let

[0018]

[0019] Then the phase consistency is defined as

[0020]

[0021] where ω o (x,y) is a weighting function so that points with smaller spectrum width have smaller weights, A so (x,y) is defined in (3), T is the noise compensation coefficient to make the phase consistency insensitive to noise, It is used to prevent the included part from being less than 0, that is, when the part is less than 0, set it to 0. ε is a constant to avoid division by 0. ΔΦ so is a phase deviation function defined as

[0022]

[0023] Its function is to make the phase consistency more accurate in locating the features.

[0024]

[0025] The phase consistency map calculated according to formula (4) can be used to detect edge points and corner points, where the corner point detection is performed by the minimum moment map m ψ The local maximum detection is performed on the edge detection by the maximum moment map M ψ Perform FAST feature detection on it to get m ψ and M ψ The definitions are as follows

[0026]

[0027]

[0028]

[0029] in

[0030]

[0031] Through feature point detection, we can get Figure 2 The position of the center pixel of each concentric circle is shown.

[0032] 2. Feature point description

[0033] The feature point description needs to reflect the geometric relationship between the feature point and its surrounding pixels. Since most of the pixels in the phase consistency map are 0, it is not suitable for constructing feature description. The feature point description is constructed on the Maximum Index Map (MIM). The construction process is as follows: Assume that in the convolution process of (2), N s Scale N o directions, for a given direction O, the log Gabor layer is defined as

[0034]

[0035] That is, each layer is N s The sum of the filtering results of the scales. Figure 3 As shown, the maximum index graph is the largest A at (x, y) o The index of the layer where (x,y) is located.

[0036] like Figure 4 As shown in , on the maximum index map, each feature point and its surrounding pixels are divided into several patches, the grayscale histogram distribution in each patch is counted, and these grayscale histograms are cascaded into a vector to obtain the feature description of the feature point.

[0037] 3. Feature point matching

[0038] Feature matching requires finding the correspondence between the feature points of each image pair. The search strategy can be similarity search or nearest neighbor search. This algorithm uses the nearest neighbor search to obtain the correspondence between the points with the same name, such as Figure 2 In fact, there are still a lot of mismatches between these matching point pairs, as shown in the matching in Figure (2) and Figure (5), so it is necessary to further screen out consistent matching relationships.

[0039] Generally, the Fast sample consensus (FSC) algorithm is used to select consistent matching relationships between matching point pairs. Suppose that N matching relationships are found on the benchmark graph and the real-time graph through the above steps. b and N r (N b =N r ) points of the same name, each time from N r Randomly select 3 points from the N points, calculate the correspondence H between them and the point pairs in the reference graph, and then apply this correspondence to all N r Points, get their image point N in the reference image r ', calculate pixel N r'Matching point N obtained in the previous step b If the distance d between them is less than the given accuracy threshold T, the point pair is considered to be a correct match. Repeat the above steps until the corresponding relationship H with the most matching point pairs is found. Summary of the Invention

[0040] Technical problems to be solved

[0041] The technical problem to be solved by the present invention is: Figure 2 It can be seen that after finding several feature points on each of the two images, it is necessary to match these points according to the search strategy. However, these feature points calculated on each image may not have a matching relationship, so that the matching point pairs obtained based on the search strategy alone contain a large number of mismatched points, as shown in Figure (5). The random sampling consistency algorithm needs to screen matching points with consistent transformation models among a large number of mismatched point pairs. Therefore, when the number of iterations is limited, the affine transformation accuracy may be poor or even an erroneous affine transformation matrix. In order to obtain a more accurate transformation matrix, the number of iterations can only be increased or even an exhaustive search can be performed. The cost of doing so is to reduce the execution efficiency of the algorithm and increase the matching time.

[0042] Technical Solution

[0043] A SAR image feature matching method based on inertial navigation error correction is characterized in that M is obtained on the reference image and the real-time image respectively. b and M r feature points, M b ≠M r , according to the search strategy, we get M b and M r There are N corresponding relationships between feature points, M b >N and M r >N, the steps are as follows:

[0044] Step 1: Estimate the position of the live image in the reference image from the N matching point pairs obtained according to the search strategy:

[0045]

[0046] Where (x ir -x ib ) is the i-th matching point in the benchmark image, (y ir -y ib ) is the i-th matching point in the real-time graph;

[0047] Step 2: Design the threshold T based on the inertial navigation error information. Assuming the inertial navigation angle error is Δθ, the threshold is set to LΔθ, where L is the width of the reference image.

[0048] Step 3: For each pair of matching points, calculate the difference δy between the y coordinates of the reference image and the real-time image. Then δy is relative to y c The size should not exceed T, that is, |δy-y c |<LΔθ, thus removing the matching points that do not meet the conditions, leaving N r matching points;

[0049] Step 4: Utilize the remaining N r Matching points calculate the affine transformation matrix H;

[0050] Step 5: Apply the transformation matrix to all N r The remaining matching points, that is, calculate the N r The image points of the matching points under the action of the affine transformation matrix H are then calculated with respect to the image points and the N r The distance between the matching points is calculated, the points whose accuracy does not meet the requirements are removed, and the remaining matching points are used as the final feature matching results.

[0051] A computer system, characterized in that it includes: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method.

[0052] A computer-readable storage medium is characterized by storing computer-executable instructions, which are used to implement the above method when executed.

[0053] Beneficial effects

[0054] The present invention provides a SAR image feature matching method based on inertial navigation error correction. The method uses matching points obtained according to a search strategy to estimate the approximate position of the real-time image within the reference image. The inertial navigation error information is then used to design a threshold for the difference between the relative position of the matching coordinates and the estimated position. Finally, this threshold is used to filter the difference between the position of the matching points and the estimated position, thereby removing a large number of abnormal mismatched points. Compared with existing feature matching algorithms, the method has the following advantages: by first eliminating some mismatched points using inertial navigation error information and then calculating the affine transformation matrix using the remaining matching points, a more accurate affine transformation matrix can be obtained with a small number of iterations, thereby improving the robustness and robustness of the algorithm. Furthermore, since fewer iterations are required, the algorithm's search time can be reduced, improving matching accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0056] Figure 1 Grayscale-based image matching algorithm;

[0057] Figure 2 Feature-based image matching algorithm;

[0058] Figure 3 MIM schematic diagram;

[0059] Figure 4 Feature description and feature vector;

[0060] Figure 5 Nearest neighbor search results containing mismatches;

[0061] Figure 6 The result after removing the mismatched points based on the inertial navigation error information. DETAILED DESCRIPTION

[0062] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only intended to illustrate 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 may be combined with each other as long as they do not conflict with each other.

[0063] Existing SAR imaging systems all rely on inertial navigation equipment to achieve motion compensation for the platform to achieve high-resolution imaging. Therefore, the results of SAR imaging can be easily transformed to the same resolution as the reference. Figure 1 When matching two images with the same north orientation, although there will still be a high number of mismatches among the resulting matching points, the inertial navigation error information can be used to constrain the positional relationships between the matching point pairs, thereby eliminating a large number of mismatches and significantly increasing the proportion of correct matching point pairs. Subsequently, a random sampling consistency algorithm is applied to these eliminated matching point pairs, allowing the calculation of the final affine transformation matrix to be completed with a reduced number of iterations.

[0064] Suppose that M is obtained on the benchmark graph and the real-time graph respectively. b and M r (M b ≠M r ) feature points, and obtain M according to the search strategy b and M r There are N corresponding relations between feature points (M b >N and M r >N), the image feature matching algorithm based on inertial navigation correction is divided into the following steps:

[0065] (1) Estimate the position of the real-time image in the reference image from the N matching point pairs obtained according to the search strategy:

[0066]

[0067] Where (x ir -x ib ) is the i-th matching point in the benchmark image, (y ir -y ib ) is the i-th matching point in the real-time graph;

[0068] The threshold T is designed based on the inertial navigation error information. Assuming the inertial navigation angle error is Δθ, the threshold is set to LΔθ, where L is the width of the reference graph.

[0069] (2) For each pair of matching points, calculate the difference δy between the y coordinates of the reference image and the real-time image, then δy is relative to y c The size should not exceed T, that is, |δy-y c |<LΔθ, thus removing the matching points that do not meet the conditions, leaving N r matching points, the results are as follows Figure 6 As shown;

[0070] (3) Using the remaining N r The affine transformation matrix H is calculated based on the matching points. At this time, since there are fewer matching points, the number of iterations required is also less.

[0071] (4) Apply the transformation matrix to all N r The remaining matching points, that is, calculate the N r The image points of the matching points under the action of the affine transformation matrix H are then calculated with respect to the image points and the N r The distance between the matching points is calculated, the points whose accuracy does not meet the requirements are removed, and the remaining matching points are used as the final feature matching results.

[0072] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present invention, and these modifications or replacements should all be included in the scope of protection of the present invention.

Claims

1. A SAR image feature matching method based on inertial navigation error correction, characterized in that Suppose that M is obtained on the benchmark graph and the real-time graph respectively. b and M r feature points, M b ≠M r , according to the search strategy, we get M b and M r There are N corresponding relationships between feature points, M b >N and M r >N, the steps are as follows: Step 1: Estimate the position of the live image in the reference image from the N matching point pairs obtained according to the search strategy: Where (x ir -x ib ) is the i-th matching point in the benchmark image, (y ir -y ib ) is the i-th matching point in the real-time graph; Step 2: Design the threshold T based on the inertial navigation error information. Assuming the inertial navigation angle error is Δθ, the threshold is set to LΔθ, where L is the width of the reference image. Step 3: For each pair of matching points, calculate the difference δy between the y coordinates of the reference image and the real-time image. Then δy is relative to y c The size should not exceed T, that is, |δy-y c |<LΔθ, thus removing the matching points that do not meet the conditions, leaving N r matching points; Step 4: Utilize the remaining N r Matching points calculate the affine transformation matrix H; Step 5: Apply the transformation matrix to all N r The remaining matching points, that is, calculate the N r The image points of the matching points under the action of the affine transformation matrix H are then calculated with respect to the image points and the N r The distance between the matching points is calculated, the points whose accuracy does not meet the requirements are removed, and the remaining matching points are used as the final feature matching results.

2. A computer system, characterized in that include: One or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the method of claim 1.

3. A computer-readable storage medium, characterized in that Computer-executable instructions are stored, and when the instructions are executed, they are used to implement the method of claim 1.

Images