A landmark-free visual measurement deformation monitoring method based on dynamic homography matrix mapping

Abstract

A method for deformation monitoring based on dynamic homography matrix mapping without marking visual measurement includes: taking natural feature images in real time with a camera, preprocessing the images and selecting the natural feature region of interest for feature recognition, obtaining the displacement results of each feature point through a feature point matching algorithm; distinguishing and marking abnormal feature points, tracing back to the nearest stable image and calculating the dynamic homography matrix, obtaining the optimal estimate of the feature points in the image through matrix mapping calculation and correcting the coordinates of the abnormal feature points; and obtaining the single displacement of the natural feature through weighted average calculation of the displacement of each feature point. The present invention realizes long-term deformation monitoring of natural features of the target to be measured without marking visual measurement, enhances the stability of natural mark recognition and tracking under the coupling of multiple environmental factors, meets the measurement accuracy and adapts to various measurement environment requirements.

Description

Technical Field

[0001] The invention relates to the technical field of computer vision, and in particular to a marker-free visual measurement deformation monitoring method based on dynamic homography matrix mapping. Background Art

[0002] With the continuous development of computer vision technology, visual measurement methods with advantages such as non-contact, long-distance, and high precision have gradually become a research hotspot in the field of structural health monitoring. However, in outdoor monitoring environments, due to the interference of environmental factors such as light changes and rainfall on visual imaging, long-term outdoor mark-free visual measurement deformation monitoring often faces the problem of unstable recognition or tracking errors, which significantly affects the accuracy of mark-free recognition. Therefore, current visual measurement methods often choose to arrange targets with obvious markings on the monitoring targets as their recognition targets. However, when the monitoring target is some special structures such as high-rise structures, ancient buildings, etc., it is difficult to install markers with obvious markings near the monitoring target due to its safety and protection requirements. Therefore, we need a mark-free target tracking method that can adapt to environmental requirements and stably identify natural marks for a long time. Summary of the invention

[0003] The present invention aims to overcome the above problems existing in the prior art and provide a marker-free visual measurement deformation monitoring method based on dynamic homography matrix mapping, so as to realize long-term marker-free visual measurement deformation monitoring of the natural features of the target to be measured, enhance the stability of natural marker recognition and tracking under the coupling of various environmental factors, meet the measurement accuracy and adapt to various measurement environment requirements.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A method for deformation monitoring by visual measurement without marking based on dynamic homography matrix mapping comprises the following steps:

[0006] (1) Select appropriate natural features at the target according to the actual working conditions, select a static site to set up the camera according to the location of the natural features, and adjust the camera parameters so that the natural features can be clearly imaged in the camera.

[0007] (2) Install filters and fill lights on the camera lens and near the camera according to the working conditions near the natural features. Calibrate the camera with a calibration board near the natural features and continuously capture images of the natural features in real time.

[0008] (3) After preprocessing the initial frame image by histogram equalization and Gaussian filtering, the region of interest is selected according to the natural feature position. The SURF algorithm response threshold is adjusted according to the feature texture, and multiple feature points and their corresponding feature descriptors are obtained through algorithm identification. A dynamic homography matrix is ​​established and its initial value is set to 1. The initial frame image is defined as the dynamic homography matrix calculation image.

[0009] (4) For the subsequent n-th frame (n≥2) image, after preprocessing such as histogram equalization and Gaussian filtering, the natural feature region of interest is selected. The number of feature points is set to adjust the SURF algorithm response threshold using the binary search method and obtain the feature point coordinates and feature descriptors through algorithm identification. The feature descriptor is used to match the feature points with the n-1-th frame image using the nearest neighbor search method (FLANN) based on the principle of minimum Euclidean distance.

[0010] (5) Based on the matching results of the feature points, the accuracy of the feature point matching between the nth frame and the n-1th frame is determined by using the two-way matching screening method and the 3σ criterion. Based on the determination results, if there are no abnormal points, the frame image is updated to a new dynamic homography matrix calculation image.

[0011] (6) If abnormal feature points appear in step (5), select the nth frame image and the dynamic homography matrix calculation image, adjust the SURF algorithm response threshold, and repeat the recognition and matching algorithm calculation to obtain a large number of feature point matching results between the two frames of images. The Progressive Consistent Sampling (PROSAC) algorithm is used to select a large number of matching feature points, and the homography matrix between the two frames of images is obtained by multiple sampling iterations. The dynamic homography matrix is ​​updated according to the calculation results.

[0012] (7) Mark the position of the feature point whose judgment result is an abnormal value, use the dynamic homography matrix mapping to obtain the coordinates of the abnormal feature point of the n-th frame image, fill it into the position of the abnormal feature point and replace the feature descriptor with the feature descriptor of the image calculated by the dynamic homography matrix.

[0013] (8) Subtract the coordinates of the feature points in the nth frame from the coordinates of the initial frame to obtain the displacement of each feature point. Use the nearest neighbor value of the feature point to evaluate its confidence, and use the weighted average algorithm to combine multiple feature displacement matrices into a single displacement result. Use the calibration coefficient to convert the natural feature pixel displacement into world coordinate displacement to obtain the real-time displacement result of the target to be measured.

[0014] Preferably, step (1) comprises: selecting a natural feature that can represent the displacement of the target to be measured according to the natural environment near the target to be measured. The principle for selecting natural features: since the subsequent homography matrix calculation needs to ensure that the feature points are in the same plane, the selection of the natural feature region of interest needs to ensure that it is in the same plane and is a rigid body motion. Select a suitable static site according to the position of the natural feature, select a suitable camera depth of field and field of view with reference to the object distance and environmental conditions, adjust the camera focal length after setting up the camera and set automatic exposure to capture clear images of the natural feature at all times.

[0015] Preferably, in step (2), considering the imaging problem of natural features in low-light environments at night, priority can be given to whether the fill-in lighting effect of urban street lamps can enable clear imaging of natural features in the camera. If there is no lighting condition in the actual environment, fill-in lights need to be set near the natural features to achieve nighttime visibility of natural features. For the problem of camera overexposure caused by direct sunlight from sunrise and sunset, a filter needs to be installed in front of the camera lens to reduce the incident light intensity so that the incident light intensity is within the adaptive range of the camera's automatic exposure, thereby ensuring clear imaging of natural features in the camera at all times. On this basis, the camera is calibrated near the natural features using a calibration plate and then the natural feature images are continuously captured in real time.

[0016] Preferably, in step (3), the initial frame image is preprocessed, and preprocessing such as using a histogram equalization algorithm to enhance image features and a Gaussian filter to eliminate image noise is performed. A suitable region of interest is selected for natural features, and the response threshold of the SURF algorithm is adjusted according to the texture complexity of the features in the image, so that the number of identified feature points is within an appropriate range to avoid excessive subsequent calculations and ensure the real-time nature of the calculation. On this basis, multiple stable feature point coordinates and their corresponding 64-bit feature descriptors are identified. In order to realize the correction of subsequent abnormal feature points, a dynamic homography matrix is ​​established and its initial value is set to 1. The initial frame image is defined as the dynamic homography matrix calculation image.

[0017] Preferably, in step (4), for the subsequent n-th frame (n≥2) of the image, the region of interest is selected after pre-processing such as using a histogram equalization algorithm to enhance image features and using a Gaussian filter to eliminate image noise. Due to the influence of all-weather lighting changes, even the same feature texture presents different image features under different light intensities, and the same algorithm response threshold may produce a completely different number of feature points. In order to ensure the real-time calculation and the accuracy of subsequent matching, the range of the number of feature points is set, and the binary search method is used to adjust the SURF algorithm response threshold and identify the feature point coordinates and their feature descriptors. On this basis, the nearest neighbor search method is used based on the 64-bit feature descriptor to match the feature points of the previous frame image on the principle of minimum Euclidean distance.

[0018] Preferably, in step (5), due to the coupling of various environmental factors such as illumination and rainfall, long-term outdoor sign-free visual measurement deformation monitoring often leads to problems of unstable recognition or abnormal matching, which significantly affects the accuracy of sign-free recognition. In order to solve this problem, a two-way matching screening method can be used to match the next frame image with the previous frame image to eliminate feature points that cannot match each other. In addition, since the natural features are in the same plane and are rigid bodies, the Euclidean distances between different matching feature points should be consistent, so the 3σ criterion is used to eliminate abnormal matching feature points based on the Euclidean distance between matching feature points. According to the screening results, if there are no abnormal points, the frame image can be regarded as a stable recognition image of the feature points and updated to a new dynamic homography matrix calculation image.

[0019] Preferably, in step (6), if abnormal feature points appear in step (5), since the previous frame image may have abnormal points, filling in the feature information of the nth frame and the n-1th frame image also has a large error, so the dynamic homography matrix calculation image is used as a stable image without abnormal points for matching. In order to improve the accuracy of subsequent homography matrix calculation, the SURF algorithm response threshold is adjusted, and the dynamic homography matrix calculation image and the nth frame image are repeatedly identified and matched to obtain a large number of feature point matching results.

[0020] Through a large amount of feature point matching data, the homography matrix model is fitted using the progressive consistent sampling algorithm. According to the feature point matching results, the feature point matching quality is estimated based on the ratio of the Euclidean distance of the feature matching points, and the formula is:

[0021] β=D1 / D2 (1)

[0022] Among them, D1 represents the minimum Euclidean distance of the feature point descriptor, and D2 represents the second minimum Euclidean distance.

[0023] Set the inlier error threshold δ, inlier number threshold ψ and maximum sampling times τ. According to the principle that the smaller the β value, the better the matching quality, all feature points are sorted in descending order according to the matching quality. Then define the growth function g(t) and select the first n′ matching feature points according to the sampling times through the growth function g(t) to establish the generation set U n ,Right now:

[0024] g(t)=min{n′:T n ′≥t} (2)

[0025] Where t represents the number of sampling times, T n ′ represents T n The samples in the subsampling come from U n The number of times, n′ represents the minimum value when the sampling number is greater than t.

[0026] From the generated set U n Select the four groups of matching feature points with the best matching quality to calculate the homography matrix H, that is:

[0027]

[0028] Among them, x m ,y m 、x n and n The pixel coordinates of the feature points of the image and the nth frame image are calculated for the dynamic homography matrix respectively.

[0029] The projection point error is calculated based on the feature point coordinates. If the error is less than the inlier error threshold δ, it is determined to be an inlier. If the number of inliers is greater than the set inlier number threshold, the number of inliers is updated. Otherwise, the number of inliers is repeatedly determined based on the number of sampling times. n And select the feature points with the second best matching quality to recalculate. Recalculate the homography matrix H and the new inliers from the updated inliers; if the sampling number is less than the maximum sampling number τ, the homography matrix model is considered to be accurately fitted, and the homography matrix H is output.

[0030] Preferably, in step (7), the feature point position whose discrimination result is an abnormal value is marked, and the dynamic homography matrix H between the image and the n-th frame image is calculated according to the dynamic homography matrix. The coordinates of the feature points of the n-th frame are mapped by calculating the dynamic homography matrix to obtain the coordinates of the feature points of the image stably identified, that is:

[0031]

[0032] in, Calculate the coordinates of the feature point numbered i in the n-th frame image corresponding to the image abnormal point in the dynamic homography matrix.

[0033] The coordinates of the feature points mapped by the dynamic homography matrix are filled into the position of the abnormal point in the nth frame. In order to achieve the matching between the feature points of the subsequent images and the feature points, the feature point descriptor of the abnormal point in the nth frame is replaced with the feature descriptor corresponding to the image calculated by the dynamic homography matrix.

[0034] Preferably, in step (8), the coordinates of all feature points in the nth frame image are subtracted from the coordinates of the feature points in the initial frame image to obtain the displacement results of each feature point. Since the errors between the matching of different feature points are different, the confidence of different feature points is evaluated by the nearest neighbor value and converted into a weight, that is:

[0035]

[0036] Among them, β i is the nearest neighbor value of the i-th feature point, ω iis the weight of the feature point with sequence number i.

[0037] The weighted average of all feature point displacements is used to obtain a single target displacement result as the natural feature pixel displacement, that is:

[0038]

[0039]

[0040] Among them, k is the number of feature points, Δx i , Δy i is the horizontal and vertical displacement of the feature point with sequence number i.

[0041] According to the calibration coefficient obtained by camera calibration, the natural feature pixel displacement is converted into displacement in world coordinates to obtain the real-time displacement result of the target to be measured.

[0042] The substantial effects of the present invention are as follows: natural feature images are captured in real time by a camera, the images are preprocessed and the natural feature regions of interest are selected for feature recognition, and the displacement results of each feature point are obtained by a feature point matching algorithm; abnormal feature points are identified and marked, the nearest stable image is traced back and a dynamic homography matrix is ​​calculated, the optimal estimation of the feature points in the image is obtained by matrix mapping calculation and the coordinates of the abnormal feature points are corrected; the displacement of each feature point is calculated by weighted average to obtain a single displacement of the natural feature.

[0043] The present invention has the following advantages:

[0044] (1) No need to install targets or other markers, expanding the scope of application of visual measurement deformation monitoring; (2) Solving the problem of abnormal matching of feature points due to environmental factors during long-term tracking of natural markers, improving the stability of visual measurement; (3) Using the confidence of feature point matching to weighted average the coordinates of each feature point of natural features, the robustness of marker-free visual measurement is improved. (4) Using the constantly changing dynamic homography matrix to calculate the image to obtain the dynamic homography matrix, the calculation error of the homography matrix is ​​reduced, and the accuracy of visual measurement is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of the method of this embodiment;

[0046] Figure 2 Schematic diagram of dynamic homography matrix mapping, in the figure:

[0047] 1,5,6,7-dynamic homography matrix calculation images; 2,3,4,8,9,10-images containing outliers; 11-dynamic homography matrix. DETAILED DESCRIPTION

[0048] The technical solution of the present invention is described below through specific embodiments in conjunction with the accompanying drawings.

[0049] (1) A method for deformation monitoring by visual measurement without marking based on dynamic homography matrix mapping, comprising the following steps: selecting a natural feature that can represent the displacement of the target to be measured according to the natural environment near the target to be measured. The principle for selecting natural features: Since the subsequent homography matrix calculation needs to ensure that the feature points are in the same plane, the selection of the region of interest of the natural feature needs to ensure that it is in the same plane and is a rigid body motion. Select a suitable static site according to the position of the natural feature, select a suitable camera depth of field and field of view with reference to the object distance and environmental conditions, adjust the camera focus after setting up the camera and set automatic exposure to capture clear images of the natural feature at all times.

[0050] (2) Considering the imaging problem of natural features in low-light environments at night, priority should be given to whether the fill-in lighting effect of urban street lamps can enable clear imaging of natural features in the camera. If there is no lighting condition in the actual environment, fill-in lights should be set near the natural features to achieve nighttime visibility of natural features. For the camera overexposure problem caused by direct sunlight from sunrise and sunset, a filter should be installed in front of the camera lens to reduce the incident light intensity so that the incident light intensity is within the adaptive range of the camera's automatic exposure, ensuring clear imaging of natural features in the camera at all times. On this basis, the camera is calibrated near the natural features using a calibration plate and then continuous real-time images of natural features are captured.

[0051] (3) Preprocess the initial frame image, using the histogram equalization algorithm to enhance image features and the Gaussian filter to eliminate image noise. Select the appropriate region of interest for the natural features, and adjust the SURF algorithm response threshold according to the texture complexity of the features in the image to keep the number of identified feature points within an appropriate range to avoid excessive subsequent calculations and ensure real-time calculations. On this basis, multiple stable feature point coordinates and their corresponding 64-bit feature descriptors are identified.

[0052] (4) For the subsequent n-th frame (n≥2) of the image, the region of interest is selected after preprocessing such as using the histogram equalization algorithm to enhance the image features and using the Gaussian filter to eliminate image noise. Due to the influence of all-weather lighting changes, even the same feature texture presents different image features under different light intensities, and the same algorithm response threshold may produce a completely different number of feature points. In order to ensure the real-time calculation and the accuracy of subsequent matching, the range of the number of feature points is set, and the binary search method is used to adjust the SURF algorithm response threshold and identify the feature point coordinates and their feature descriptors. On this basis, the nearest neighbor search method is used based on the 64-bit feature descriptor to match the feature points of the previous frame image with the principle of minimum Euclidean distance.

[0053] (5) Due to the coupling of various environmental factors such as illumination and rainfall, long-term outdoor sign-free visual measurement deformation monitoring often encounters problems of unstable recognition or abnormal matching, which significantly affects the accuracy of sign-free recognition. In order to solve this problem, a two-way matching screening method can be used to match the next frame image with the previous frame image to eliminate feature points that cannot match each other. In addition, since the natural features are in the same plane and are rigid bodies, the Euclidean distances between different matching feature points should be consistent, so the 3σ criterion is used to eliminate abnormal matching feature points based on the Euclidean distance between matching feature points. According to the screening results, if there are no abnormal points, the frame image can be regarded as a stable recognition image of feature points and updated to a new dynamic homography matrix calculation image.

[0054] (6) If abnormal feature points appear in step (5), since the previous frame image may have abnormal points, filling the image feature information of the nth frame and the n-1th frame image also has a large error, so the dynamic homography matrix calculation image is used as a stable image without abnormal points for matching. In order to improve the accuracy of subsequent homography matrix calculation, the SURF algorithm response threshold is adjusted, and the dynamic homography matrix calculation image and the nth frame image are repeatedly identified and matched to obtain a large number of feature point matching results.

[0055] Through a large amount of feature point matching data, the homography matrix model is fitted using the progressive consistent sampling algorithm. According to the feature point matching results, the feature point matching quality is estimated based on the ratio of the Euclidean distance of the feature matching points, and the formula is:

[0056] β=D1 / D2 (1)

[0057] Among them, D1 represents the minimum Euclidean distance of the feature point descriptor, and D2 represents the second minimum Euclidean distance.

[0058] Set the inlier error threshold δ, inlier number threshold ψ and maximum sampling times τ. According to the principle that the smaller the β value, the better the matching quality, all feature points are sorted in descending order according to the matching quality. Then define the growth function g(t) and select the first n′ matching feature points according to the sampling times through the growth function g(t) to establish the generation set U n ,Right now:

[0059] g(t)=min{n′:T n ′≥t} (2)

[0060] Where t represents the number of sampling times, T n ′ represents T n The samples in the subsampling come from U n The number of times, n′ represents the minimum value when the sampling number is greater than t.

[0061] From the generated set Un Select the four groups of matching feature points with the best matching quality to calculate the homography matrix H, that is:

[0062]

[0063] Among them, x m ,y m 、x n and n The pixel coordinates of the feature points of the image and the nth frame image are calculated for the dynamic homography matrix respectively.

[0064] The projection point error is calculated based on the feature point coordinates. If the error is less than the inlier error threshold δ, it is determined to be an inlier. If the number of inliers is greater than the set inlier number threshold, the number of inliers is updated. Otherwise, the number of inliers is repeatedly determined based on the number of sampling times. n And select the feature points with the second best matching quality to recalculate. Recalculate the homography matrix H and the new inliers from the updated inliers; if the sampling number is less than the maximum sampling number τ, the homography matrix model is considered to be accurately fitted, and the homography matrix H is output.

[0065] (7) Mark the position of the feature point that is identified as an outlier, and calculate the dynamic homography matrix H between the image and the n-th frame image based on the dynamic homography matrix. The coordinates of the feature points of the n-th frame are mapped by calculating the dynamic homography matrix to obtain the coordinates of the feature points of the image stably identified, that is:

[0066]

[0067] in, Calculate the coordinates of the feature point numbered i in the n-th frame image corresponding to the image abnormal point in the dynamic homography matrix.

[0068] Fill the feature point coordinates obtained by dynamic homography matrix mapping into the abnormal point position of the nth frame, and the result is as follows Figure 2 As shown, the image containing abnormal feature points is corrected using the dynamic homography matrix. In order to achieve the matching between the feature points of subsequent images and the feature points, the feature point descriptor of the abnormal point of the nth frame is replaced with the feature descriptor corresponding to the image calculated by the dynamic homography matrix.

[0069] (8) Subtract the coordinates of all feature points in the nth frame image from the coordinates of the feature points in the initial frame image to obtain the displacement results of each feature point. Since the errors between matching different feature points are different, the nearest neighbor value is used to evaluate the confidence of different feature points and convert it into a weight, that is:

[0070]

[0071] Among them, β i is the nearest neighbor value of the i-th feature point, ω iis the weight of the feature point with sequence number i.

[0072] The weighted average of all feature point displacements is used to obtain a single target displacement result as the natural feature pixel displacement, that is:

[0073]

[0074]

[0075] Among them, k is the number of feature points, Δx i , Δy i is the horizontal and vertical displacement of the feature point with sequence number i.

[0076] According to the calibration coefficient obtained by camera calibration, the natural feature pixel displacement is converted into displacement in world coordinates to obtain the real-time displacement result of the target to be measured.

[0077] This embodiment uses a dynamic homography matrix to solve the problem of abnormal identification and matching of natural features due to environmental factors in long-term displacement identification, and realizes long-term stable identification and tracking of natural features. Since the errors between different feature points obtained by natural feature identification are also different, the confidence of each feature point is evaluated by the nearest neighbor value, and the feature points are weighted averaged according to the confidence to obtain the final displacement result of the natural feature, thereby improving the accuracy of mark-free visual measurement.

[0078] The homography matrix is ​​a matrix that describes the projection relationship of images at different times. The dynamic homography matrix is ​​updated by calculating the feature points that correctly match the image at the current moment through a stable and high-quality dynamic homography matrix. The coordinates of the abnormal feature points contained in the image at the current moment are corrected through dynamic homography matrix mapping to obtain the displacement of each feature point of the natural feature of the image at the current moment for a long time. This method not only solves the problem of abnormal matching feature points caused by environmental factors during the long-term tracking of natural marks, improves the stability of visual measurement, but also realizes the mark-free visual measurement deformation monitoring without installing targets and other markers, expanding the scope of application of visual measurement deformation monitoring. In addition, the dynamic homography matrix calculated by using the constantly changing dynamic homography matrix to calculate the image reduces the error of the homography matrix calculation and improves the accuracy of visual measurement.

[0079] The above embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention.

Claims

1. A method for deformation monitoring by visual measurement without marking based on dynamic homography matrix mapping, comprising the following steps: (1) Select appropriate natural features at the target according to the actual working conditions, select a static site to set up the camera according to the location of the natural features, and adjust the camera parameters so that the natural features can be clearly imaged in the camera; (2) Install filters and fill lights on the camera lens and near the camera according to the working conditions near the natural features; calibrate the camera using a calibration plate near the natural features and continuously capture images of the natural features in real time; (3) After preprocessing the initial frame image by histogram equalization, Gaussian filtering, etc., the region of interest is selected according to the natural feature position; the SURF algorithm response threshold is adjusted according to the feature texture and multiple feature points and their corresponding feature descriptors are obtained through algorithm identification; a dynamic homography matrix is ​​established and its initial value is set to 1; the initial frame image is defined as the dynamic homography matrix calculation image; (4) For the subsequent n-th frame (n≥2) of the image, after preprocessing such as histogram equalization and Gaussian filtering, the region of interest of natural features is selected; The number range of feature points is set, and the binary search method is used to adjust the response threshold of the SURF algorithm, and the coordinates of the feature points and their feature descriptors are obtained through algorithm identification; Using the feature descriptor, the nearest neighbor search method (FLANN) is used to match the feature points with the n-1 frame image based on the principle of minimum Euclidean distance; (5) According to the matching results of the feature points, the accuracy of the feature point matching between the nth frame and the n-1th frame is determined by using the two-way matching screening method and the 3σ criterion; based on the determination results, if there are no abnormal points, the frame image is updated to a new dynamic homography matrix calculation image; (6) If abnormal feature points appear in step (5), select the nth frame image and the dynamic homography matrix calculation image, adjust the SURF algorithm response threshold, repeat the recognition and matching algorithm calculation to obtain a large number of feature point matching results between the two frames of images; use the Progressive Consistent Sampling (PROSAC) algorithm through a large number of matching feature points, give priority to selecting high-quality matching points and calculate the homography matrix between the two frames of images through multiple sampling iterations, and update the dynamic homography matrix according to the calculation results; specifically including: if abnormal feature points appear in step (5), since the previous frame image may have abnormal points, there is a large error in filling the feature information of the nth frame and the n-1th frame image, so the dynamic homography matrix calculation image is used as a stable image without abnormal points for matching; in order to improve the accuracy of subsequent homography matrix calculation, adjust the SURF algorithm response threshold, repeat the recognition and matching algorithm calculation to obtain the dynamic homography matrix calculation image and the nth frame image to obtain a large number of feature point matching results; Through a large amount of feature point matching data, the homography matrix model is fitted using the progressive consistent sampling algorithm; according to the feature point matching results, the feature point matching quality is estimated based on the ratio β of the Euclidean distance of the feature matching points. The formula is: β=D1 / D2 (1) Among them, D1 represents the minimum value of the Euclidean distance of the feature point descriptor, and D2 represents the second minimum value of the Euclidean distance; Set the inlier error threshold δ, the inlier number threshold ψ and the maximum sampling number τ; according to the principle that the smaller the β value, the better the matching quality, all feature points are sorted in descending order according to the matching quality, and then the growth function g(t) is defined and the growth function g(t) is used according to the sampling number. Select the first n′ matching feature points to establish the generation set U n ,Right now: g(t)=min{n′:T n ′≥t} (2) Where t represents the number of sampling times, T n ′ represents T n The samples in the subsampling come from U n The number of times, n′ represents the minimum value when the sampling number is greater than t; From the generated set U n Select the four groups of matching feature points with the best matching quality to calculate the homography matrix H, that is: Among them, x m ,y m 、x n and n Calculate the pixel coordinates of the feature points of the image and the nth frame image for the dynamic homography matrix respectively; The projection point error is calculated based on the feature point coordinates. If the error is less than the inlier error threshold δ, it is determined to be an inlier. If the number of inliers is greater than the set inlier number threshold, the number of inliers is updated. Otherwise, the number of inliers is repeatedly determined based on the number of sampling times. n And select the feature points with the second best matching quality to recalculate; recalculate the homography matrix H and the new internal points based on the updated internal points; if the sampling number is less than the maximum sampling number τ, it is considered that the homography matrix model is accurately fitted, and the homography matrix H is output; (7) Mark the position of the feature point whose judgment result is an abnormal value, use the dynamic homography matrix mapping to obtain the coordinates of the abnormal feature point of the n-th frame image, fill it into the abnormal feature point position and replace the feature descriptor with the feature descriptor of the image calculated by the dynamic homography matrix; (8) The coordinates of the feature points in the nth frame are subtracted from the coordinates of the initial frame to obtain the displacement of each feature point. The confidence of the feature point is evaluated using the nearest neighbor value. The weighted average algorithm is used to combine the multiple feature displacement matrices into a single displacement result. The natural feature pixel displacement is converted into world coordinate displacement through the calibration coefficient to obtain the real-time displacement result of the target to be measured.

2. The method for deformation monitoring by visual measurement without marking based on dynamic homography matrix mapping according to claim 1, characterized in that: Step (1) includes: selecting a natural feature that can represent the displacement of the target to be measured according to the natural environment near the target to be measured; the principle for selecting natural features: since the subsequent homography matrix calculation needs to ensure that the feature points are in the same plane, the selection of the natural feature area of ​​interest needs to ensure that it is in the same plane and is a rigid body motion; selecting a suitable static site according to the position of the natural feature, selecting a suitable camera depth of field and field of view with reference to the object distance and environmental conditions, and adjusting the camera focal length after setting up the camera and setting automatic exposure to capture clear images of the natural feature at all times.

3. The method for deformation monitoring by visual measurement without marking based on dynamic homography matrix mapping as claimed in claim 1, characterized in that: In step (2), considering the imaging problem of natural features in low-light environment at night, priority can be given to whether the fill light effect of urban street lamps can make the natural features clearly imaged in the camera. If there is no lighting condition in the actual environment, fill lights need to be set near the natural features to achieve nighttime visibility of the natural features. For the camera overexposure problem caused by direct sunlight from sunrise and sunset, a filter needs to be installed in front of the camera lens to reduce the incident light intensity so that the incident light intensity is within the adaptive range of the camera's automatic exposure, thereby ensuring clear imaging of the natural features in the camera at all times. On this basis, the camera is calibrated near the natural features using a calibration plate and then continuous real-time shooting of natural feature images is carried out.

4. The method for deformation monitoring by visual measurement without marking based on dynamic homography matrix mapping according to claim 1, characterized in that: In step (3), the initial frame image is preprocessed, and the image features are enhanced by the histogram equalization algorithm and the image noise is eliminated by Gaussian filtering; a suitable region of interest is selected for the natural features, and the response threshold of the SURF algorithm is adjusted according to the texture complexity of the features in the image, so that the number of identified feature points is within an appropriate range to avoid excessive subsequent calculations and ensure the real-time nature of the calculations; on this basis, multiple stable feature point coordinates and their corresponding 64-bit feature descriptors are identified; in order to realize the correction of subsequent abnormal feature points, a dynamic homography matrix is ​​established and its initial value is set to 1; the initial frame image is defined as the dynamic homography matrix calculation image.

5. The method for deformation monitoring by visual measurement without marking based on dynamic homography matrix mapping according to claim 1, characterized in that: In step (4), for the subsequent n-th frame (n≥2) image, the region of interest is selected after pre-processing such as using a histogram equalization algorithm to enhance image features and using a Gaussian filter to eliminate image noise; due to the influence of all-weather illumination changes, even the same feature texture presents different image features under different light intensities, and the same algorithm response threshold may produce a completely different number of feature points; in order to ensure the real-time calculation and the accuracy of subsequent matching, the range of the number of feature points is set, and the binary search method is used to adjust the SURF algorithm response threshold and identify the feature point coordinates and their feature descriptors; on this basis, the nearest neighbor search method is used based on the 64-bit feature descriptor to match the feature points of the previous frame image with the principle of minimum Euclidean distance.

6. The method for deformation monitoring by visual measurement without marking based on dynamic homography matrix mapping according to claim 1, characterized in that: In step (5), a two-way matching screening method is used to match the next frame image with the previous frame image to eliminate feature points that cannot match each other. Since the natural features are in the same plane and are rigid bodies, the Euclidean distances between different matching feature points should be consistent. Therefore, the 3σ criterion is used to eliminate abnormal matching feature points according to the Euclidean distances between matching feature points. According to the screening results, if there are no abnormal points, the frame image can be regarded as a stable recognition image of feature points and updated to a new dynamic homography matrix calculation image.

7. The method for deformation monitoring by visual measurement without marking based on dynamic homography matrix mapping according to claim 1, characterized in that: In step (7), the position of the feature point whose judgment result is an abnormal value is marked, and the dynamic homography matrix H between the dynamic homography matrix calculation image and the n-th frame image is calculated according to the calculation; the coordinates of the feature points of the n-th frame are mapped by the dynamic homography matrix calculation image stable recognition, that is: in, Calculate the coordinates of the feature point numbered i in the nth frame image corresponding to the image abnormal point in the dynamic homography matrix; The coordinates of the feature points obtained by mapping the dynamic homography matrix are filled into the position of the abnormal point in the nth frame; in order to achieve the matching between the feature points of subsequent images and the feature points, the feature point descriptor of the abnormal point in the nth frame is replaced by the feature descriptor corresponding to the image calculated by the dynamic homography matrix.

8. The method for deformation monitoring by visual measurement without marking based on dynamic homography matrix mapping as claimed in claim 1, characterized in that: In step (8), the coordinates of all feature points in the nth frame image are subtracted from the coordinates of the feature points in the initial frame image to obtain the displacement results of each feature point; since the errors between the matching of different feature points are different, the nearest neighbor value is used to evaluate the confidence of different feature points and convert it into a weight, that is: Among them, β i is the nearest neighbor value of the i-th feature point, ω i is the weight of the feature point with sequence number i; The weighted average of all feature point displacements is used to obtain a single target displacement result as the natural feature pixel displacement, that is: Among them, k is the number of feature points, Δx i , Δy i is the horizontal and vertical displacement of the feature point with sequence number i; According to the calibration coefficient obtained by camera calibration, the natural feature pixel displacement is converted into displacement in world coordinates to obtain the real-time displacement result of the target to be measured.

Images