A tunnel apparent intelligent inspection robot image rapid splicing method

Abstract

The application discloses a kind of tunnel apparent intelligent inspection robot's image fast splicing method, it is related to the field of apparent image splicing technology under tunnel scene, comprising: real-time acquisition tunnel lining multichannel apparent image of tunnel apparent intelligent inspection robot collection;Identify the hole in apparent image and construction joint, and correct the construction joint identified;Using the hole identified and the construction joint corrected, longitudinally align each channel apparent image, and fine-tune the longitudinal alignment result;By ring splicing algorithm, the overlapping redundant pixels between each channel apparent image are merged;According to segment, output ring spliced segment image.The hole identification method of the application is more real-time, and can cope with complex hole environment at the same time.The ring alignment method of the application is better for the characteristics of current acquisition equipment and tunnel.The construction joint identification result correction method of the application is not disturbed by abnormal misjudgment construction joint, by projecting the identification result to two-dimensional plane for filtering, so as to filter false positive result.

Description

Technical Field

[0001] This invention relates to the field of appearance image stitching technology in tunnel scenarios, and more specifically to a rapid image stitching method for a tunnel appearance intelligent inspection robot. It is applicable to the image recognition of tunnel entrances in tunnel scenarios, the image feature point recognition and matching algorithm applicable to tunnel scenarios (for texture circumferential alignment), and a construction joint recognition algorithm with an error correction mechanism (construction joint recognition is used for texture longitudinal alignment). The key point is the application of the construction joint recognition algorithm and the filtering algorithm in the direction of correcting the construction joint recognition results. Background Technology

[0002] Tunnel lining, as a key component supporting the stability of the tunnel structure and preventing deformation of the surrounding rock and groundwater seepage, has an appearance condition (such as cracks, leaks, spalling, and defects) that directly affects the operational safety and service life of the tunnel. Therefore, regular, efficient, and accurate appearance inspection of tunnel lining is a core aspect of infrastructure maintenance.

[0003] Traditional tunnel lining inspection mainly relies on manual inspection and localized photography using handheld devices. This method suffers from low efficiency, strong subjectivity, incomplete coverage, high risk, and difficulty in digital archiving and historical comparison. In recent years, with the development of mobile platforms (such as inspection vehicles and robots) equipped with imaging systems, image-based automated inspection technology has become the mainstream. Among these, multi-channel imaging systems are widely used to acquire 360° circumferential surface images of the tunnel in a single operation. This system typically deploys multiple cameras (e.g., 4, 6, or 8) evenly along the circumference of the mobile platform, triggering synchronously as the platform moves to continuously acquire images of the inner surface of the tunnel lining, thereby generating massive amounts of sequential image data.

[0004] However, to transform these multi-channel, sequential local images into continuous panoramic images that can be used for panoramic analysis, automatic defect identification and quantification, efficient and accurate image stitching technology is an indispensable key step.

[0005] In the rapid image stitching of intelligent tunnel surface inspection robots, tunnel entrance positioning and image alignment are the core steps.

[0006] 1. Tunnel entrance positioning The shape of the opening is like Figure 6 As shown in the diagram of a certain tunnel entrance, there are generally two existing methods for locating tunnel entrances: Method 1: One approach is to adaptively binarize the image of the tunnel entrance, calculate the proportion of black and white areas, and then set an empirical threshold for the black and white ratio of the entrance. Entrances with a black and white ratio exceeding this threshold are considered entrances; those exceeding the threshold are not. The problem with this algorithm is that various types of obstacles may exist outside the tunnel entrance, such as… Figure 7As shown in image 1 (not an opening), the algorithm misidentifies this type of image as an opening; simultaneously, as... Figure 8 As shown in image 2 (not an entrance), this type of safety escape tunnel within a tunnel would be identified as an entrance.

[0007] Method 2: Another method is to use a deep neural network to perform binary classification on images of tunnel entrances and non-tunnel entrances. The problem with this algorithm is that the inspection equipment requires high real-time performance. If a neural network is used for identification, it is difficult to meet its requirement of rapid positioning. At the same time, since the amount of data on tunnel entrances is small, and a tunnel only has two entrances (tunnel entrance and tunnel exit), the neural network is very prone to overfitting when the data samples are too few.

[0008] 2. Image alignment Existing image alignment algorithms typically follow the following algorithm flow: Figure 9 As shown, the algorithm first finds the feature points of the image pair to be aligned using a feature extraction algorithm, then finds the matching feature points in the two images using a matching algorithm, and finally solves the homography matrix of image 2 transformed into image 1. The problem with this algorithm in the field of tunnel image alignment is that, due to the small overlap between cameras in tunnel image acquisition devices and the fact that most areas of the tunnel wall have similar texture features and no obvious feature points, this algorithm is very prone to errors in feature matching environments. Summary of the Invention

[0009] To overcome the shortcomings of existing technologies, this invention discloses a rapid image stitching method for a tunnel appearance intelligent inspection robot. The purpose of this invention is to address the problems existing in current tunnel entrance location methods and image alignment algorithms. This invention proposes a mathematically based entrance location algorithm and automatically finds the decision plane between entrances and non-entrances using an SVM decision-maker. This avoids the inability of manually selected thresholds to handle various changing scenarios. Simultaneously, it increases the number of features, avoiding the overly simplistic use of the proportion of black and white areas in the image, which is insufficient for complex scenarios. This method can more easily identify entrances and non-entrances. Furthermore, this invention proposes a feature matching algorithm suitable for tunnel scenarios, which can achieve better image alignment.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A rapid image stitching method for a tunnel appearance intelligent inspection robot includes the following steps: S1. Real-time acquisition of multi-channel surface images of tunnel lining collected by intelligent tunnel appearance inspection robot; S2. Perform feature augmentation processing on the appearance image and use an SVM classifier to identify the holes in the appearance image after feature augmentation. S3. Identify construction joints in the acquired appearance image, and then process them sequentially through coordinate transformation, Gaussian filtering, and threshold filtering, combined with a window fusion algorithm to correct the identified construction joints; S4. Using the identified openings and corrected construction joints, align the appearance images of each channel longitudinally and fine-tune the longitudinal alignment results. S5. Merge redundant pixels that overlap between the apparent images of each channel using a circumferential stitching algorithm; S6. Output the segmented image that has been stitched together in a circumferential direction according to the segments.

[0011] I. Fast Positioning Algorithm for Cave Entrance Preferably, in step S2, the step of performing feature augmentation processing on the appearance image and using an SVM classifier to identify holes in the appearance image after feature augmentation includes: S211. Perform grayscale stretching on the apparent image; S212. Calculate the grayscale histogram of the image after grayscale stretching; S213. Perform adaptive binarization on the grayscale stretched image, and perform morphological processing on the binarized image. S214. Flip the grayscale histogram and calculate the gradient of the flipped histogram. S215. In the binarized image after morphological processing, calculate the mean gray value of the black region, the mean gray value of the white region, and the proportion of the black region. S216. Input the histogram gradient, the mean gray value of the black area of ​​the image, the mean gray value of the white area of ​​the image, and the proportion of the black area of ​​the image into the SVM classifier, find the best decision plane for holes and non-holes under the four features, and determine whether it is a hole.

[0012] Preferably, in step S213, the morphological processing includes: first performing a morphological opening operation to eliminate small foreign objects, and then performing a morphological closing operation to close the unclosed image regions.

[0013] Preferably, in step S214, flipping the grayscale histogram and calculating the gradient of the flipped histogram includes: moving the histogram along... x =127.5 Flip, and transform the gradient histogram according to the following formula. The domain after transformation is... ;

[0014] in, The grayscale value after histogram conversion is The total number of pixels; The grayscale value is The number of pixels; Grayscale value It belongs to [0, 255].

[0015] Preferably, in step S216, finding the optimal decision plane for openings and non-openings under four features includes: taking the extracted histogram gradient, the mean gray value of the black area of ​​the image, the mean gray value of the white area of ​​the image, and the proportion of the black area of ​​the image as feature vectors and inputting them into the SVM classifier. After learning and analysis, the SVM classifier obtains the corresponding decision function, i.e. the optimal decision plane, and then uses the decision function to identify and classify the appearance image to be detected.

[0016] II. Construction joint identification algorithm and construction joint identification result correction algorithm Preferably, in step S3, the construction joint in the acquired appearance image includes: The apparent image is input into a neural network model, which then identifies the coordinate information of the construction joint. x i , y i , w i , h i , score i );in, x i Let x be the x-coordinate of the i-th construction joint. y i Let be the ordinate of the i-th construction joint. w i Let be the width of the i-th construction joint. h i Let be the height of the i-th construction joint. score i Let be the confidence score of the i-th construction joint; The neural network model sorts all identified construction joints according to their confidence scores in the construction joint coordinate information, and takes the construction joint with the highest confidence score as the output.

[0017] Preferably, the confidence score of the construction joint is calculated as follows:

[0018] in, , The sigmoid function maps output values ​​to the interval (0, 1). For the output value of category c, The predicted output value for IoU. and These are all output results predicted by neural network models.

[0019] Preferably, in step S3, the correction of the identified construction joint includes: projecting the identified construction joint information into a two-dimensional coordinate system, filtering the construction joint in the two-dimensional coordinate system using Gaussian filtering, and then outputting the filtered construction joint after threshold filtering and window fusion algorithm.

[0020] Preferably, in step S3, during the correction of the identified construction joint, the construction joint information includes coordinate information, which is: the coordinate information in the camera ID of the first... The first picture OK; The step of projecting the identified construction joint information into a two-dimensional coordinate system includes: converting the construction joint coordinates into absolute coordinates according to the following construction joint coordinate transformation formula:

[0021] in, Coordinates of the construction joint; The index of the image in the camera; It is a translation matrix; This represents the row number of the image where the construction joint is located.

[0022] III. Circular Alignment Method Preferably, in step S5, the circumferential stitching algorithm includes: Read in the first circumferential stitched image and the second circumferential stitched image; Calculate the feature points of the overlapping region of the first and second circumferential stitched images, and perform feature point matching in the overlapping region. When matching feature points, sort them by the horizontal pixel coordinates of the images as close as possible. Based on the feature point matching results, the homography matrix of the second circumferential stitched image transformed into the first circumferential stitched image is calculated, and the homography matrix is ​​used to perform circumferential stitching of the images.

[0023] In step S4, the sorting of images by their horizontal pixel coordinates as close as possible includes: using image processing algorithms to extract corresponding feature points in the first and second circumferential images, and then using the feature points in the first circumferential image as a basis to calculate the feature points that are closest to the feature points in the second circumferential image as the final matching feature points.

[0024] The beneficial effects of this invention are: 1. This invention proposes a mathematical hole location algorithm and uses an SVM decision-maker to automatically find decision planes for holes and non-holes, thereby avoiding the inability of manually selected thresholds to cope with various changing scenarios. At the same time, it increases the number of features and avoids the overly simplistic use of the proportion of black and white areas in the image, which is unable to cope with complex scenarios. This method can more easily identify non-holes and holes.

[0025] 2. In this invention, the tunnel opening identification method offers higher real-time performance and can handle complex tunnel opening environments. The circumferential alignment method of this invention provides better stitching results tailored to the characteristics of current acquisition equipment and tunnels.

[0026] 3. The construction joint identification result correction method of the present invention is not affected by abnormal false detection of construction joints. By projecting the identification result onto a two-dimensional plane for filtering, false alarm results are filtered out. Attached Figure Description

[0027] Figure 1 This is a general framework diagram of the rapid image stitching method for the intelligent tunnel appearance inspection robot of the present invention; Figure 2 This is a schematic diagram of the hole recognition algorithm of the present invention; Figure 3 This is a schematic diagram of the construction joint identification algorithm of the present invention; Figure 4 This is the construction joint correction algorithm flow of the present invention; Figure 5 This is a schematic diagram of the circumferential splicing feature matching of the present invention; Figure 6 The shape of a certain cave entrance; Figure 7 Image 1 is the non-hole image; Figure 8 Image 2 is the non-hole image; Figure 9 This describes the existing image matching algorithm process. Detailed Implementation

[0028] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with the embodiments and accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention.

[0029] Example 1 A rapid image stitching method for an intelligent tunnel appearance inspection robot mainly involves image recognition of tunnel entrances, image feature point recognition and matching algorithms applicable to tunnel scenarios, and a construction joint recognition algorithm with an error correction mechanism. The key points are the application of the construction joint recognition algorithm and the filtering algorithm in the direction of construction joint correction. The overall block diagram is shown in Figure 1, which uses eight cameras to capture images of the tunnel, thus ensuring a complete record of the tunnel.

[0030] The overall framework process is as follows: 1. Acquire tunnel images collected by the intelligent tunnel appearance inspection robot.

[0031] 2. Identify tunnel entrances and construction joints in tunnel images collected by the intelligent tunnel appearance inspection robot.

[0032] 3. The software performs vertical alignment of the channel data based on the identified openings and construction joints (using the opening and construction joint of the first channel as a reference, it calculates the distance between the construction joints of other channels and the construction joint of the first channel, and then performs vertical translation according to this distance to complete the alignment).

[0033] 4. Manually fine-tune the results of the software's automatic vertical alignment.

[0034] 5. Redundant pixels that overlap between channels are merged using a circumferential stitching algorithm.

[0035] 6. Finally, output the circumferentially spliced ​​segment images according to segments (i.e., the area between two construction joints is considered as a segment).

[0036] Example 2 This embodiment further elaborates on the rapid hole location algorithm based on the above embodiments.

[0037] Algorithm such as Figure 2 As shown, the algorithm achieves a recognition speed of 10ms on a platform with an i5-9400f CPU and Windows 10. This solves the problem of poor real-time performance in the existing solution using deep neural networks. Furthermore, by adding features (such as histogram features, the mean grayscale value of the black area after morphological processing of the original image, the mean grayscale value of the white area after morphological processing of the original image, and the black-and-white ratio of the image after morphological processing), and by automatically finding the decision plane of the hole and non-hole based on the above features using SVM, the algorithm avoids the problem that the manually selected black-and-white ratio threshold in the existing solution cannot handle complex scenes.

[0038] like Figure 2 As shown, the specific algorithm flow is as follows: 1. Input image.

[0039] 2. Perform grayscale stretching on the image. Since the tunnel image is generally dark, grayscale stretching can significantly improve the contrast between the inside and outside of the tunnel entrance.

[0040] 3. Calculate the grayscale histogram of the stretched image for subsequent feature calculation.

[0041] 4. Perform adaptive binarization on the image after gray-scale stretching. At this time, for the image data of the hole without foreign objects, the inside and outside of the hole can already be segmented. However, if there are foreign objects, interference will occur, so step 5 is carried out.

[0042] 5. For the binarized image, first perform morphological opening operation to eliminate small foreign objects, and then perform closing operation to close the unclosed image area.

[0043] 6. Calculate the features of the output results of the above five steps. First, flip the histogram along x = 127.5, that is, convert the gradient histogram according to the following formula. After conversion, the domain . The idea of this formula is that after the gray-scale stretching of the hole image, the outside of the hole is almost all black, and the inside of the hole is almost all white. At this time, the histogram will show Figure 2 the peaks at x = 0 and x = 255 shown in the histogram in 3, and the middle interval (0 < x < 255) is almost 0. However, due to the interference of foreign objects, there will be some values in the middle. But the overall histogram of the hole image will show a "U" shape, while non-hole areas may show a "one" shape. At this time, after flipping the histogram, if it is a hole, the gradient of its histogram must be less than 0, while non-hole areas may be greater than 0 or both greater than 0 and less than 0 exist. Therefore, calculating the gradient of the flipped histogram will be a powerful feature to distinguish hole images and non-hole images, that is, the gradient feature of the flipped histogram.

[0044]

[0045] Among them, is the total number of pixels with the gray value after histogram conversion; is the number of pixels with the gray value ; is the gray value, belongs to [0, 255].

[0046] 7. In addition to the histogram feature, according to the binary image after morphological processing, calculate the gray mean value of this area in the image after gray-scale stretching as a feature based on the range of its black area, because the average gray value of the black area of the hole is often very low, while the average gray value of, for example, a lawn (which may also be processed into a black area) is relatively high. Therefore, the gray value feature of the black area helps to distinguish holes and non-holes.

[0047] 8. Similar to step 7, according to the binary image after morphological processing, calculate the gray mean value of the white area in the image after gray-scale stretching as a feature based on the range of its white area, that is, the gray value feature of the white area.

[0048] 9. In the binary image after statistical morphological processing, the proportion of black areas to the entire image is used as a feature, namely the black area proportion feature.

[0049] 10. Input the four features—the gradient of the flipped histogram, the mean gray value of the black region, the mean gray value of the white region, and the proportion of the black region—into an SVM (Support Vector Machine) classifier to automatically find the optimal decision plane for holes and non-holes based on these four features. Details are as follows: The extracted histogram gradient, the mean gray value of the black area of ​​the image, the mean gray value of the white area of ​​the image, and the proportion of the black area of ​​the image are used as feature vectors and input into the SVM classifier. After learning and analysis, the SVM classifier obtains the corresponding decision function, i.e. the optimal decision plane, and then uses the decision function to identify and classify the appearance image to be detected.

[0050] 11. During real-time detection, four features of the image are calculated, and then an SVM is used to determine whether it is a hole.

[0051] Example 3 This embodiment further elaborates on the construction joint identification algorithm and the construction joint identification result correction algorithm based on the above embodiments.

[0052] Construction joint recognition algorithm such as Figure 3 As shown, after reading the image, it is input into the neural network model, and then the model outputs the coordinate information of the construction joint identified by the model, where the score represents the confidence score. Due to the special nature of the acquisition equipment capturing tunnel construction joints, if a construction joint exists in a single acquired image, there will be at most one construction joint. Therefore, the algorithm sorts all identified construction joints according to their confidence scores and takes the construction joint with the highest confidence score as the output.

[0053] Specifically, the coordinate information of the construction joint is ( x i , y i , w i , h i , score i );in, x i Let x be the x-coordinate of the i-th construction joint. y i Let be the ordinate of the i-th construction joint. w i Let be the width of the i-th construction joint. h i Let be the height of the i-th construction joint. score i Let be the confidence score of the i-th construction joint.

[0054] The confidence score for construction joints is calculated as follows:

[0055] in, , The sigmoid function maps output values ​​to the interval (0, 1). For the output value of category c, The predicted output value for IoU. and These are all output results predicted by neural network models.

[0056] After all construction joints in the images are identified, a large amount of construction joint information will be obtained (such as...). Figure 4 As shown in (1), Includes its coordinate information, which are the first in the camera ID. The first picture Okay. Projecting a large amount of construction joint information onto a two-dimensional coordinate system, where the construction joint coordinates are converted to absolute coordinates using the following formula, we can see... Figure 4 (2) The light blue construction seams are falsely identified construction seams. This is because there are theoretically 8 construction seams at the construction seam, but some images from the camera may not be detected, so the number is less than 8. The light blue ones are isolated points, so they must be noise points (i.e. falsely reported construction seams). Then, a Gaussian filter (with a kernel size of 3) is used to filter them. It can be seen that the falsely identified construction seams are greatly suppressed after filtering. Finally, the filtered construction seams are output after threshold filtering and window fusion algorithm.

[0057]

[0058] in, Coordinates of the construction joint; The index of the image in the camera; It is a translation matrix; This represents the row number of the image where the construction joint is located.

[0059] Example 4 This embodiment further elaborates on the circumferential alignment method based on the above embodiments.

[0060] Because there are some overlapping areas between the cameras, feature point matching in the overlapping areas can be used to complete the circumferential stitching. However, due to the significant differences in brightness between the cameras and the lack of obvious features on most of the tunnel walls, feature matching errors are very likely to occur. Currently, the main difficulty in circumferential alignment lies in the fact that cameras 1-4 and 5-8 did not capture images simultaneously. Due to the frame dropping characteristic of line scan cameras, misalignment occurs during circumferential stitching. Therefore, this method designs a matching algorithm specifically for the characteristics of current devices. That is, when matching feature points, instead of sorting them by similarity, it sorts them by the horizontal pixel coordinates of the image being as close as possible. Figure 9 As shown, the difference between the image matching algorithm and the one used on-site lies in the matching of feature points. The basis for sorting the images by their horizontal pixel coordinates as close as possible is that a line of data acquired by the line scan camera can be saved accurately, while frame loss occurs during the vertical acquisition process.

[0061] Specifically, circumferential alignment methods include: Read in the first circumferential stitched image and the second circumferential stitched image; Calculate the feature points of the overlapping region of the first and second circumferential stitched images, and perform feature point matching in the overlapping region. When matching feature points, the feature points are sorted by the horizontal pixel coordinates of the images as close as possible. Specifically, this includes: using image processing algorithms to extract the corresponding feature points in the first and second circumferential images, and then using the feature points in the first circumferential image as a basis, calculating the feature points that are closest to the feature points in the second circumferential image as the final matching feature points. Based on the feature point matching results, the homography matrix of the second circumferential stitched image transformed into the first circumferential stitched image is calculated, and the homography matrix is ​​used to perform circumferential stitching of the images.

[0062] like Figure 5 As shown, (a) is an image taken by the left camera, (b) is an image taken by the right camera, and (c) is an image of the areas that may overlap, which are then used for feature point detection and matching.

[0063] The embodiments of the present invention have been described in detail above, but the present invention is not limited to the described embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalents or substitutions are all included within the scope defined by the claims of the present invention.

Claims

1. A rapid image stitching method for a tunnel appearance intelligent inspection robot, characterized in that, Includes the following steps: Real-time acquisition of multi-channel surface images of tunnel lining collected by intelligent tunnel appearance inspection robot; The appearance image is subjected to feature augmentation processing, and an SVM classifier is used to identify the holes in the appearance image after feature augmentation. The construction joints in the acquired appearance image are identified, and then the identified construction joints are corrected by coordinate transformation, Gaussian filtering, threshold filtering, and window fusion algorithm. Using the identified openings and corrected construction joints, the visual images of each passage are longitudinally aligned, and the longitudinal alignment results are fine-tuned. The redundant pixels overlapping between the appearance images of each channel are merged using a circumferential stitching algorithm; Output the segmented image in a circumferential stitching pattern according to the segments.

2. The rapid image stitching method for a tunnel appearance intelligent inspection robot as described in claim 1, characterized in that, The step of performing feature augmentation processing on the appearance image and using an SVM classifier to identify holes in the appearance image after feature augmentation includes: Perform grayscale stretching on the apparent image; Calculate the grayscale histogram of the image after grayscale stretching; Adaptive binarization is performed on the grayscale stretched image, and morphological processing is then applied to the binarized image. Flip the grayscale histogram and calculate the gradient of the flipped histogram; In the binarized image after morphological processing, calculate the mean gray value of the black region, the mean gray value of the white region, and the proportion of the black region. The histogram gradient, the mean gray value of the black area of ​​the image, the mean gray value of the white area of ​​the image, and the proportion of the black area of ​​the image are input into the SVM classifier to find the best decision plane for holes and non-holes under the four features, and determine whether it is a hole.

3. The rapid image stitching method for a tunnel appearance intelligent inspection robot as described in claim 2, characterized in that, The morphological processing includes: first performing a morphological opening operation to eliminate small foreign objects, and then performing a morphological closing operation to close the unclosed image regions.

4. The rapid image stitching method for a tunnel appearance intelligent inspection robot as described in claim 2, characterized in that, The step of flipping the grayscale histogram and calculating the gradient of the flipped histogram includes: moving the histogram along... x =127.5 Flip, and transform the gradient histogram according to the following formula. The domain after transformation is... ; in, The grayscale value after histogram conversion is The total number of pixels; The grayscale value is The number of pixels; This is the grayscale value.

5. The rapid image stitching method for a tunnel appearance intelligent inspection robot as described in claim 1, characterized in that, The construction joints in the acquired appearance images include: The apparent image is input into a neural network model, which then identifies the coordinate information of the construction joint. x i , y i , w i , h i , score i );in, x i Let x be the x-coordinate of the i-th construction joint. y i Let be the ordinate of the i-th construction joint. w i Let be the width of the i-th construction joint. h i Let be the height of the i-th construction joint. score i Let be the confidence score of the i-th construction joint; The neural network model sorts all identified construction joints according to their confidence scores in the construction joint coordinate information, and takes the construction joint with the highest confidence score as the output.

6. The rapid image stitching method for a tunnel appearance intelligent inspection robot as described in claim 5, characterized in that, The confidence score of the construction joint is calculated as follows: in, , The sigmoid function maps output values ​​to the interval (0, 1). For the output value of category c, The predicted output value for IoU. and These are all output results predicted by neural network models.

7. The rapid image stitching method for a tunnel appearance intelligent inspection robot as described in claim 1, characterized in that, The correction of the identified construction joints includes: projecting the identified construction joint information into a two-dimensional coordinate system, filtering the construction joints in the two-dimensional coordinate system using Gaussian filtering, and then outputting the filtered construction joints after threshold filtering and window fusion algorithm.

8. The rapid image stitching method for a tunnel appearance intelligent inspection robot as described in claim 7, characterized in that, In the process of correcting the identified construction joints, the construction joint information includes coordinate information, which is: the coordinate information in the camera ID. The first picture OK; The step of projecting the identified construction joint information into a two-dimensional coordinate system includes: converting the construction joint coordinates into absolute coordinates according to the following construction joint coordinate transformation formula: in, Coordinates of the construction joint; The index of the image in the camera; It is a translation matrix; This represents the row number of the image where the construction joint is located.

9. The rapid image stitching method for a tunnel appearance intelligent inspection robot as described in claim 1, characterized in that, The circumferential stitching algorithm includes: Read in the first circumferential stitched image and the second circumferential stitched image; Calculate the feature points of the overlapping region of the first and second circumferential stitched images, and perform feature point matching in the overlapping region. When matching feature points, sort them by the horizontal pixel coordinates of the images as close as possible. Based on the feature point matching results, the homography matrix of the second circumferential stitched image transformed into the first circumferential stitched image is calculated, and the homography matrix is ​​used to perform circumferential stitching of the images.

10. The rapid image stitching method for a tunnel appearance intelligent inspection robot as described in claim 9, characterized in that, The sorting of images by horizontal pixel coordinates as close as possible includes: using image processing algorithms to extract corresponding feature points in the first and second circumferential images, and then using the feature points in the first circumferential image as a basis to calculate the feature points that are closest to the feature points in the second circumferential image as the final matching feature points.

Images