CN110363706A - A large-area bridge deck image stitching method - Google Patents

Abstract

The invention discloses a kind of large area bridge floor image split-joint methods.With the development of Computer Vision Detection Technique, gradually appears and image detection is applied in bridge machinery engineering practice.But since pontic space is larger, according to long-range shooting sampling, it will receive the limitation of resolution of video camera, cause to be unable to get satisfied detection accuracy.The present invention is as follows: 1, acquiring the image for being detected bridge floor one by one, obtain bridge floor image collection.Later, image preprocessing is carried out.2, image registration.3, image co-registration.The present invention completes the automation non-destructive testing of bridge defect feature instead of human eye by Image Acquisition and processing technique, has very important realistic meaning to the research of the bridge floor damage detection technology under complicated landform environment.On the one hand working security is enhanced, on the other hand improves operation mobility and flexibility.Invention realizes the fidelity splicing of large area bridge floor image, improves image mosaic precision and efficiency.

Description

A large-area bridge deck image stitching method

technical field

The invention belongs to the technical field of image detection, and in particular relates to a large-area bridge deck image stitching method.

Background technique

Bridges are important transportation hubs. After long-term sun, rain, and load operations, the internal stress will be transmitted to some weak parts along the bridge structure, resulting in cracks and other disease characteristics on the surface of the structure. Due to the disease characteristics of the bridge surface, the outside air and harmful media can easily penetrate into the concrete to generate carbonate through chemical reaction, resulting in the reduction of the alkalinity environment of the steel bar, and the rust is more likely to occur after the surface purification film is damaged. It will also aggravate shrinkage cracking and cause serious harm to the safe use of concrete bridges. Therefore, in order to ensure the service life and safety performance of the beam structure, it is necessary to detect and treat the disease characteristics of the bridge surface in time.

In order to detect the characteristics of bridge deck diseases in time and take remedial measures to eliminate potential safety hazards, manual inspection and manual marking are usually adopted. However, the working environment of bridge bottom surface inspection is often dangerous. This inspection method has poor mobility, high risk and low efficiency. Moreover, the bridge deck damage is manually measured by experienced inspectors and recorded with the naked eye, which is subject to a certain degree of subjectivity. The detection accuracy is more dependent on the experience and knowledge of experts, and experience lacks objectivity in quantitative analysis.

With the development of computer vision detection technology, the application of image detection in bridge detection engineering practice has gradually emerged. However, due to the large space of the bridge body, if remote shooting sampling is used, it will be limited by the resolution of the camera, resulting in unsatisfactory detection accuracy. Therefore, it is necessary to sample the bridge surface in consecutive groups in a short distance, and perform large-area image stitching on the collected sample images, and the stitching effect will directly affect the detection accuracy of the bridge deck disease characteristics. In order to reduce the cumulative error as much as possible, improve the stitching accuracy of bridge deck images, and realize the global display of real bridge inspection surface information on the basis of satisfying wide viewing angle and high resolution.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a large-area bridge deck image stitching method.

The concrete steps of the present invention are as follows:

Step 1. Collect the images of the detected bridge deck one by one to obtain a set of bridge deck images. After that, the luminance component information of each bridge deck image in the bridge deck image set is extracted and counted, and the luminance component information of each bridge deck image is equalized respectively. Then, each bridge deck image is transformed into the frequency domain by Fourier transform, and the phase information of the normalized cross-power spectrum in the phase correlation algorithm is used to obtain the translation parameters between the images, and the pre-estimation of the overlapping area between adjacent images is completed. .

Step 2. Image registration

First, SIFT feature points are extracted in the overlapping area between adjacent images. Then, the SIFT feature points are filtered by the adaptive contrast threshold method, and the feature descriptors composed of matching point pairs are obtained. And the RANSAC algorithm is used to calculate the projection transformation matrix between adjacent images.

The adaptive contrast threshold method is as follows:

(1) Set the lower limit of the number of feature points N _min =200, the upper limit N _max =300, and the contrast threshold T _c =T ₀ . T ₀ is the initial threshold, which ranges from 0.02 to 0.04.

(2) Perform feature point detection, and count the number N of feature points whose contrast is higher than _Tc .

(3) If N _min ≤N≤N _max , the feature points with contrast higher than T _c are included in the initial matching point set, the feature points with contrast lower than the threshold T _c are eliminated, and step (5) is directly entered. Otherwise, go to step (4).

(4) If N<N _min , reduce the contrast threshold T _c to the value of the original value and perform step (3). If N>N _max , increase the contrast threshold to twice the original value, and execute step (3).

(5) The erroneous feature points in the initial matching point set are eliminated by the nearest neighbor ratio and the next nearest neighbor method, and feature descriptors are generated. The feature descriptor contains multiple matching point pairs composed of paired feature points, and the distance and direction information between each matching point pair.

Step 3. Image fusion

Firstly, according to the projection transformation matrix between adjacent images, projective transformation is performed on the corresponding bridge deck images. Then, the RGB three-color channel of each adjacent bridge deck image is weighted and smoothly transitioned by the fade-in and fade-out fusion algorithm, and the bridge deck stitched image is obtained.

In the fade-in and fade-out fusion algorithm, the fade-in and fade-out weighting formula of each fusion point pixel value I(x, y) in the overlapping area of adjacent images is as follows:

Among them, I ₁ (x, y) and I ₂ (x, y) are the pixel values of the corresponding fusion points of the two adjacent bridge deck images in the overlapping area, respectively. d ₁ and d ₂ are the gradient weight factors of the two adjacent bridge deck images at the corresponding fusion points, respectively. and x ₁ and x ₂ are the abscissas of the borders on both sides of the overlapping area, respectively. x is the abscissa of the corresponding fusion point. t is the grayscale difference threshold at the corresponding fusion point in the overlapping area of two adjacent images.

Preferably, the method for collecting images in step 1 is as follows:

(1) After calculating the internal parameter matrix of the CCD camera by Zhang Zhengyou's plane calibration method, the radial distortion coefficient is obtained by the least square method.

(2) Install the CCD camera calibrated in step 1-1 on the bridge inspection platform, and collect images of the complete bridge deck according to the preset shooting trajectory. The preset shooting trajectory is S-shaped.

(3) Perform image calibration on each bridge deck image collected in step 1-2 according to the internal parameter matrix and distortion coefficient of the CCD camera obtained in step 1-1.

Preferably, in step 2, the process of solving the projection transformation matrix by the RANSAC algorithm is as follows:

(1) Construct an initial sample set S with each matching point pair in the feature descriptor. Count the Euclidean distances between each matching point pair in the initial sample set S, and sort them from small to large.

(2) Take the first 85% matching point pairs of the sequence obtained in step (1) to construct a new sample set S'.

(3) Randomly select 4 sets of matching point pairs from the new sample set S' to form an interior point set S _i , and calculate the H _i of the interior point set S _i of the matrix model, and enter step (4).

(4) The remaining matching point pairs in the new sample set S' are tested for adaptability of the matrix model _Hi . If there is a matching point whose inspection error is smaller than the error threshold, the matching point pair whose inspection error is smaller than the threshold is added to the interior point set S _i , and step (5) is executed. Otherwise, discard the matrix model H _i and execute (3) again.

(5) If the number of elements in the interior point set Si is greater than the specified threshold, it is considered that a reasonable parameter model is obtained _{, and the matrix model H i is} recalculated for the updated interior point set Si _, and the LM algorithm is used to minimize the cost function. Otherwise, discard the matrix model H _i and perform step (3) again.

(6) Repeat steps (3) to (5) for l times, where l is the maximum number of iterations. Afterwards, compare the interior point set Si obtained in l iterations, take the interior point set Si with the largest number of elements as the final interior point set _, and take the calculated matrix model H _i as the difference between adjacent bridge deck images _. Projection transformation matrix.

Preferably, in step 3, the specific steps of projection transformation are as follows:

(1) According to the transitivity of the projection transformation matrix between adjacent images, the first bridge deck image of each row is used as the reference image of the corresponding row for stitching. Perform transfer transformation on the transformation matrix H _ii-1 between each adjacent bridge deck image to obtain the transfer transformation matrix H _i1 between each bridge deck image and the reference image. Then, the corresponding bridge deck images are mapped into the reference plane coordinate system through each transformation matrix H _i1 to complete the image splicing and fusion between adjacent images in the horizontal direction to form multiple horizontal panoramic images Image _i with wide viewing angles.

(2) Stitching the first horizontal panoramic image Image ₁ obtained in step (1) as a reference panoramic image. Perform transfer transformation on the transformation matrix T _jj-1 between each horizontal panorama image to obtain a transfer transformation matrix T _j1 between each horizontal panorama image Image _i and the reference panorama image. Then, the corresponding horizontal panoramic images are respectively mapped into the reference plane coordinate system through each transfer transformation matrix T _j1 to complete the image splicing and fusion between adjacent horizontal panoramic images in the vertical direction to form the final bridge deck panoramic image.

The beneficial effects that the present invention has are:

1. The present invention uses the image acquisition and processing technology to replace the human eye to complete the automatic nondestructive detection of bridge damage characteristics, which has very important practical significance for the research of bridge deck damage detection technology under complex terrain environment. On the one hand, the construction safety is enhanced, and on the other hand, the operation mobility and flexibility are improved.

2. The present invention proposes an improved multi-group adjacent bridge deck image registration algorithm in order to extract the disease characteristic data of bridge deck images more completely and accurately, aiming at the problems of large computational load and insufficient precision of traditional image registration algorithms. , to achieve the fidelity stitching of large-area bridge deck images, improve the accuracy and efficiency of image registration, lay a foundation for the subsequent image detection of bridge disease characteristics, and also provide a technical reference for image stitching detection in other fields.

3. Aiming at the special detection environment of the bridge surface, the present invention proposes an improved fade-in and fade-out image fusion algorithm, and introduces the grayscale difference threshold of the adjacent bridge deck images, which can effectively suppress the influence of irrelevant noise of the bridge deck images and preserve the maximum extent. The detailed feature information of the bridge deck disease can improve the signal-to-noise ratio of the stitched images on the basis of realizing the fidelity fusion of multiple groups of bridge deck images.

4. The present invention improves the anti-interference and stability of the bridge deck image stitching algorithm under complex background, and has better robustness. Ensure the accuracy and precision of subsequent disease feature data extraction.

5. According to the working environment of bridge surface image collection, the present invention designs multiple groups of bridge deck image stitching algorithms, and conducts reliability analysis and research on key calculations, which has high algorithm innovation and reference value for bridge inspection projects.

Description of drawings

Fig. 1 is the large-area bridge deck image stitching flow chart in the present invention;

2 is a schematic diagram of a bridge deck image acquisition track in the present invention;

Fig. 3 is a flow chart of the stitching processing of multiple bridge images in the present invention;

Fig. 4 is the self-adaptive contrast threshold calculation flow chart in the present invention;

5 is a schematic diagram of a line-by-line image stitching strategy in the present invention;

6 is a schematic diagram of a column-by-column image stitching strategy in the present invention;

7a and 7b are schematic diagrams of weighted fusion between adjacent images in the present invention.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings.

As shown in Figure 1, a large-area bridge deck image stitching method, the specific steps are as follows:

Step 1. Splicing preprocessing

1-1. Using the Zhang Zhengyou plane calibration method, use the CCD camera to sample images of the calibration board from different angles, calculate the internal parameter matrix of the CCD camera through the detected calibration board checkerboard corner coordinates, and obtain the radial distortion coefficient through the least squares method. .

1-2. Install the CCD camera calibrated in step 1-1 on the bridge inspection platform, and collect multiple sets of images of the complete bridge deck according to the preset shooting trajectory. The preset shooting track is shown in Figure 2, which is S-shaped.

1-3. Perform image calibration on each bridge deck image collected in step 1-2 according to the internal parameter matrix and distortion coefficient of the CCD camera obtained in step 1-1, so as to eliminate the distortion effect caused by lens distortion.

1-4. Image preprocessing

First pre-estimate the size of the panorama image. The size is valued according to the resolution and quantity of the images to be spliced, and the invalid area is removed after the splicing is completed; then, by extracting and counting the brightness component information of all the bridge deck images to be spliced, and separately performing the brightness component information of each bridge deck image. Equalization to eliminate the influence of brightness difference caused by uneven illumination; finally, each bridge deck image is transformed into the frequency domain through Fourier transform, and the phase information of the normalized cross-power spectrum in the phase correlation algorithm is used to obtain the image between the images. to complete the pre-estimation of the overlapping area between adjacent images.

The phase correlation algorithm uses Fourier transform to first transform the image to be spliced into the frequency domain, and then calculates the translation parameters of the two images through the normalized cross-power spectrum to obtain a two-dimensional impulse function: the peak value of the two-dimensional impulse function. The size reflects the content correlation between adjacent bridge deck images. A value of 1 indicates that the two images are identical, and a value of 0 indicates that they are completely different. There are changes caused by perspective transformation and position movement between adjacent images collected by bridge image detection equipment. Although the energy of the impulse function will be dispersed from a single peak to many small peaks, the translation parameters corresponding to the maximum peak position will remain relatively stable. Therefore, the translation amount obtained by the phase correlation algorithm can roughly obtain the overlapping area between the images to be spliced, and the algorithm is insensitive to changes in illumination brightness, and the detected maximum correlation peak has good robustness and stability.

Step 2. Image registration

First, SIFT feature points are extracted in the overlapping area between adjacent images to reduce a large number of unnecessary feature point detection calculations and improve the detection efficiency of SIFT feature points; The number of points is controlled within a reasonable range to screen out stable feature point sets; and the improved RANSAC algorithm (random sampling consistency algorithm) is used to calculate the projection transformation matrix H between adjacent images.

The adaptive contrast threshold method is as follows:

After determining the overlapping area (Δx, Δy) between two adjacent bridge deck images, SIFT feature point detection is performed only for the overlapping area. Since the feature points with low contrast are more sensitive to the background noise of the bridge deck, the contrast threshold is set to filter out a stable feature point set, which is denoted as C.

In the prior art, the contrast of each SIFT feature point is calculated by Gaussian difference Taylor expansion, and a fixed contrast threshold is set to retain feature points higher than the contrast threshold as stable feature points. However, the above-mentioned contrast threshold T _c is a fixed value, and generally ranges from 0.02 to 0.04. However, in the crack image detection of different concrete bridges, the candidate feature point sets detected by SIFT are quite different. Some bridge surfaces are relatively smooth and clean, the collected image digital signals are relatively smooth, and the scale space factor σ is small, resulting in the detection of There are fewer feature points, but it may not be able to meet the number of feature point matching requirements, which affects the final stitching accuracy. (Contrast thresholding step used in traditional stitching algorithms). The present invention sets a variable contrast threshold to ensure that the number of detected SIFT feature points is controlled within a reasonable range. Several sets of experimental verifications show that the feature points of bridge crack image detection can be kept between 200 and 300 to meet better stitching accuracy.

As shown in Figure 4, the method for determining the contrast threshold T _c in the present invention is as follows:

(1) Set the lower limit of the number of feature points N _min =200, the upper limit N _max =300, and the contrast threshold T _c =T ₀ ; T ₀ is the initial threshold, which is 0.02 to 0.04.

(2) Perform feature point detection, and count the number N of feature points whose contrast is higher than _Tc .

(3) If N _min ≤N≤N _max , the feature points with contrast higher than T _c are included in the initial matching point set, the feature points with contrast lower than the threshold T _c are eliminated, and directly enter step (5); otherwise, execute Step (4).

(4) If N<N _min , reduce the contrast threshold T _c to the value of the original value And execute step (3); if N>N _max , increase the contrast threshold to twice the original value, and execute step (3).

(5) The erroneous feature points in the initial matching point set are eliminated by the nearest neighbor ratio and the next nearest neighbor method, and feature descriptors are generated. The feature descriptor contains multiple matching point pairs composed of paired feature points, and the distance and direction information between each matching point pair.

After the feature points between the overlapping areas of adjacent images are matched, enough matching point pairs are selected, and the transformation matrix between the bridge crack sequence images is solved through the matching point pairs, so as to complete the large-scale bridge deck image stitching. In order to further improve the efficiency and accuracy of image registration, the RANSAC algorithm is improved.

The process of the improved RANSAC algorithm to solve the projection transformation matrix H is as follows:

(1) Construct an initial sample set S with each matching point pair in the feature descriptor. Count the Euclidean distances between matching point pairs in the initial sample set S, and sort them from small to large;

(2) taking the first 85% matching point pairs of the sequence obtained in step (1) to construct a new sample set S';

(3) randomly select 4 groups of matching point pairs from the new sample set S' to form an interior point set S _i , and calculate the H _i of the interior point set S _i of the matrix model, and enter step (4);

(4) The remaining matching point pairs in the new sample set S′ are tested for adaptability of the matrix model _Hi ; if there are matching points whose test error is less than the error threshold, the matching point pairs whose test error is less than the threshold will be added to the interior point set S _i , and execute step (5); otherwise, discard the matrix model H _i and execute step (3) again.

(5) If the number of elements in the interior point set Si is greater than the specified threshold, it is considered that a reasonable parameter model is obtained _{, the matrix model H i is} recalculated for the updated interior point set Si _, and the LM algorithm is used to minimize the cost function; Otherwise, discard the matrix model H _i and perform step (3) again.

(6) Repeat steps (3) to (5) for l times, where l is the maximum number of iterations. Afterwards, compare the interior point set Si obtained in l iterations, take the interior point set Si with the largest number of elements as the final interior point set _, and take the calculated matrix model H _i as the difference between adjacent bridge deck images _. Projection transformation matrix H.

The improved RANSAC algorithm calculates the Euclidean distance between all matching point pairs and sorts them, which not only reduces the sample set data of the point pairs to be matched, but also increases the proportion of in-office points in the sample set, and reduces the iteration of the projection transformation matrix. The number of refinements to improve the matching accuracy of the bridge deck image. According to the characteristic that the smaller the distance between the image feature point pairs, the higher the matching similarity, the Euclidean distance between all feature point pairs is calculated here and sorted in ascending order for screening. According to the statistics of the stitching test results of multiple groups of bridge deck images, after the initial matching of the feature point pairs in the overlapping area, the successful matching rate of the initial sample set S can reach more than 85%, then the first 85% of the feature point pairs in the sequence are used to construct a new sample set S'. After sample data screening, the sample set S' contains enough matching point pairs, which not only increases the proportion of intra-office points in the sample set, but also greatly reduces the number of iterations of the transformation matrix parameter model H.

Step 3. Image fusion

Firstly, according to the projection transformation matrix between adjacent images, the corresponding bridge deck image is subjected to projective transformation; then the RGB three-color channels of each adjacent bridge deck image are weighted and smoothly transitioned by the fade-in and fade-out fusion algorithm to obtain the bridge deck mosaic. image.

The specific steps of projection transformation are as follows:

(1) As shown in Figure 5, according to the transitivity of the projection transformation matrix between adjacent images, the first bridge deck image of each row is used as the reference image of the corresponding row, and the splicing is performed according to the image row splicing strategy. The transformation matrix H _ii-1 between the adjacent bridge deck images is transferred and transformed to obtain the transfer transformation matrix H _i1 between each bridge deck image and the reference image _; It is mapped into the reference plane coordinate system to complete the image splicing and fusion between adjacent images in the horizontal direction to form multiple horizontal panoramic images Image _i with a wide viewing angle.

The transformation formula of each transfer matrix is as follows:

H ₂₁ =H ₂₁

H ₃₁ =H ₃₂ ×H ₂₁

H _n1 =H _nn-1 ×H _n-1n-2 ×…×H ₂₁

Among them, H _ii-1 is the transformation matrix between the i-1 th bridge deck image in the same row and the i th bridge deck image, and its value is calculated in step 2-2; H _i1 is the first image in the same row The transformation matrix between the bridge deck image and the ith bridge deck image; n is the number of images on the same row.

(2) As shown in FIG. 6 , the first horizontal panoramic image Image ₁ obtained in step (1) is used as a reference panoramic image, and is stitched according to the image column stitching strategy. Transfer transformation is carried out to the transformation matrix T _jj-1 between each horizontal panorama image, obtains the transfer transformation matrix T _j1 between each horizontal panorama image Image _i and the reference panoramic image _; The horizontal panoramic images are respectively mapped into the reference plane coordinate system to complete the image splicing and fusion between adjacent horizontal panoramic images in the vertical direction. In the image splicing and fusion, the transfer matrix transformation formula refers to the description in step (1) to form the final image. Panoramic image of bridge deck.

In this embodiment, the fade-in and fade-out fusion algorithm is improved, and the details are as follows.

After image registration, in order to further eliminate the interference of bridge deck image stitching on crack detection processing, the pixel values of adjacent bridge deck images are usually weighted and averaged, as shown in Figure 7b, the pixels in the overlapping area are stitched to both sides. The distance of the line is used as the basis for judging the fusion weight.

However, due to the change of the image acquisition position, the reflected light from the bridge surface may cause the gray value of individual pixels in the overlapping area to jump. a threshold t. Calculate the grayscale difference corresponding to the two original images of the overlapped target pixel. If the difference is less than the threshold, it means that the pixel does not show significant difference in the original bridge deck image, and its weighted average can be directly taken as the point. On the contrary, it means that the image to be spliced has a sudden change of light and dark at this pixel position, and the pixel value with a larger weight before smoothing should be taken as the fusion pixel value at this point.

In the fade-in and fade-out fusion algorithm, the fade-in and fade-out weighting formula of each fusion pixel value I(x, y) in the overlapping area of adjacent images is as follows:

Among them, I ₁ (x, y) and I ₂ (x, y) are the pixel values of the corresponding fusion points of the two adjacent bridge deck images in the overlapping area, as shown in Figure 7a. d ₁ and d ₂ are the gradient weight factors of the two adjacent bridge deck images at the corresponding fusion points respectively; as shown in Figure 7b, and x ₁ and x ₂ are the abscissas of the borders on both sides of the overlapping area, respectively; x is the abscissa of the corresponding fusion point; t is the grayscale difference threshold at the corresponding fusion point in the overlapping area of two adjacent images.

Claims (4)

1. a large-area bridge deck image stitching method is characterized in that: step 1, collect the images of the detected bridge deck one by one to obtain a set of bridge deck images; after that, extract and count the number of images of each bridge deck in the set of bridge deck images. The luminance component information of each bridge deck image is equalized respectively; then each bridge deck image is transformed into the frequency domain through Fourier transform, and the phase of the normalized cross-power spectrum in the phase correlation algorithm is used. The information obtains the translation parameters between the images, and completes the pre-estimation of the overlapping area between adjacent images; Step 2. Image registration First, extract SIFT feature points in the overlapping area between adjacent images; then filter SIFT feature points by adaptive contrast threshold method to obtain feature descriptors composed of matching point pairs; and use RANSAC algorithm to calculate the difference between adjacent images. Projection transformation matrix; The adaptive contrast threshold method is as follows: (1) Set the lower limit of the number of feature points N _min = 200, the upper limit N _max = 300, and the contrast threshold T _c =T ₀ ; T ₀ is the initial threshold, which ranges from 0.02 to 0.04; (2) Perform feature point detection, and count the number N of feature points whose contrast is higher than _Tc ; (3) If N _min ≤N≤N _max , the feature points with contrast higher than T _c are included in the initial matching point set, the feature points with contrast lower than the threshold T _c are eliminated, and directly enter step (5); otherwise, execute step (4); (4) If N<N _min , reduce the contrast threshold T _c to the value of the original value And execute step (3); if N>N _max , increase the contrast threshold to 2 times the original value, and execute step (3); (5) Eliminate erroneous feature points in the initial matching point set by the nearest neighbor ratio and second nearest neighbor method, and generate a feature descriptor; the feature descriptor contains multiple matching point pairs composed of paired feature points, and each matching point pair. distance and direction information between; Step 3. Image fusion Firstly, according to the projection transformation matrix between adjacent images, the corresponding bridge deck image is subjected to projective transformation; then the RGB three-color channels of each adjacent bridge deck image are weighted and smoothly transitioned by the fade-in and fade-out fusion algorithm to obtain the bridge deck mosaic. image; In the fade-in and fade-out fusion algorithm, the fade-in and fade-out weighting formula of each fusion point pixel value I(x, y) in the overlapping area of adjacent images is as follows: Among them, I ₁ (x, y) and I ₂ (x, y) are the pixel values of the corresponding fusion points of the adjacent two bridge deck images in the overlapping area; d ₁ and d ₂ are the adjacent two The gradient weight factor of the bridge deck image at the corresponding fusion point; and x ₁ and x ₂ are the abscissas of the borders on both sides of the overlapping area, respectively; x is the abscissa of the corresponding fusion point; t is the grayscale difference threshold at the corresponding fusion point in the overlapping area of two adjacent images. 2. a kind of large-area bridge deck image stitching method according to claim 1 is characterized in that: the method for collecting images in step 1 is as follows: (1) After calculating the internal parameter matrix of the CCD camera by Zhang Zhengyou's plane calibration method, the radial distortion coefficient is obtained by the least square method; (2) The CCD camera calibrated in step 1-1 is placed on the bridge inspection platform, and the image acquisition of the complete bridge deck is carried out according to the preset shooting track; the preset shooting track is S-shaped; (3) Perform image calibration on each bridge deck image collected in step 1-2 according to the internal parameter matrix and distortion coefficient of the CCD camera obtained in step 1-1. 3. a kind of large-area bridge deck image stitching method according to claim 1, is characterized in that: in step 2, the process flow that RANSAC algorithm solves projection transformation matrix is as follows: (1) Construct an initial sample set S with each matching point pair in the feature descriptor; count the Euclidean distance between each matching point pair in the initial sample set S, and sort them from small to large; (2) taking the first 85% matching point pairs of the sequence obtained in step (1) to construct a new sample set S'; (3) randomly select 4 groups of matching point pairs from the new sample set S' to form an interior point set S _i , and calculate the H _i of the interior point set S _i of the matrix model, and enter step (4); (4) The remaining matching point pairs in the new sample set S′ are tested for adaptability of the matrix model _Hi ; if there are matching points whose test error is less than the error threshold, the matching point pairs whose test error is less than the threshold will be added to the interior point set S _i , and execute step (5); otherwise, discard the matrix model H _i and execute step (3) again; (5) If the number of elements in the interior point set Si is greater than the specified threshold, it is considered that a reasonable parameter model is obtained _{, the matrix model H i is} recalculated for the updated interior point set Si _, and the LM algorithm is used to minimize the cost function; Otherwise, discard the matrix model H _i and perform step (3) again; (6) Repeat steps (3) to (5) for l times, where l is the maximum number of iterations; after that, compare the interior point sets Si obtained in the l _iterations , and take the interior point set Si with the largest number of _elements as the final and take the calculated matrix model _Hi as the projection transformation matrix between adjacent bridge deck images. 4. a kind of large-area bridge deck image stitching method according to claim 1, is characterized in that: in step 3, the concrete steps of projection transformation are as follows: (1) According to the transitivity of the projection transformation matrix between the adjacent images, the first bridge deck image of each row is used as the reference image of the corresponding row for splicing; the transformation matrix H _{ii- 1.} Carry out transfer transformation to obtain the transfer transformation matrix H _i1 between each bridge deck image and the reference image; and then map the corresponding bridge deck image into the reference plane coordinate system through each transformation matrix H _i1 to complete the horizontal direction. Image splicing and fusion between adjacent images to form multiple horizontal panoramic images Image _i with wide viewing angles; (2) splicing the first horizontal panorama image Image ₁ obtained in step (1) as a reference panorama image; transfer the transformation matrix T _jj-1 between each horizontal panorama image to obtain each horizontal panorama image Image _i and the transfer transformation matrix T _j1 between the reference panoramic image; then through each transfer transformation matrix T _j1 , the corresponding horizontal panoramic image is respectively mapped into the reference plane coordinate system, to complete the vertical direction between each adjacent horizontal panoramic image The images are stitched and fused to form the final panoramic image of the bridge deck.