A tunnel disease detection method and system based on a UAV

Abstract

The embodiment of the present application provides a kind of tunnel disease detection method and system based on unmanned aerial vehicle, belong to the field of defect light measurement.The method comprises: control unmanned aerial vehicle flies along tunnel and collects the multi-channel image data of structure surface, and based on ambient light information adjusts flight attitude to carry out image light compensation;Based on the structure anchor point atlas constructed by historical acquisition image, combined with the multi-channel image data after light compensation of unmanned aerial vehicle flight pose information, the image alignment of each target anchor point is carried out;Disease area is identified in the image of alignment, the disease characteristics of each target anchor point this time are extracted, and the disease of each target anchor point this time is updated into the time sequence characteristic data sequence of each target anchor point corresponding;Based on the time sequence characteristic data sequence of each target anchor point after updating, the disease evolution trend of tunnel is analyzed using bayesian variable point detection.The present application scheme improves the alignment accuracy, time sequence comparability and risk judgment ability of tunnel disease detection as a whole.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of defect light measurement, in particular to a tunnel disease detection method based on a UAV and a tunnel disease detection system based on a UAV. BACKGROUND

[0002] Tunnel structures are in complex environments such as high humidity, low light, or uneven lighting for a long time, and are prone to diseases such as cracks, corrosion, and water seepage. In order to ensure the safety of the tunnel structure, image acquisition and image processing combined disease detection means have been widely used. With the rapid development of UAV platforms, UAVs equipped with optical imaging devices have gradually become an important tool for tunnel disease detection. Through flight tasks to complete image acquisition in the tunnel, and image analysis and recognition, remote structure evaluation can be achieved efficiently and with low risk.

[0003] However, the existing detection scheme based on UAV images still faces two technical challenges in practical application. First, due to the complex lighting conditions inside the tunnel, uneven distribution of lighting devices or alternating strong and weak sections, there are phenomena such as overexposure, overexposure, or shadow superposition in the image, which seriously affect the clarity and recognizability of the disease features (such as crack texture, edge profile, etc.) in the image. Some existing methods attempt to compensate through histogram equalization, contrast enhancement, and other means of single-frame images, but in the face of continuous scenes and dynamic lighting changes in tunnels, it is often difficult to effectively adapt to all areas, resulting in insufficient accuracy of image feature extraction.

[0004] Second, the disease development process has obvious time sequence, and engineering maintenance personnel often need to compare the evolution trend of the disease at the same position based on images from different detection periods to realize state judgment and risk warning. However, due to environmental disturbances, positioning errors, and other factors in different flight rounds of the UAV, it is difficult to ensure that the target area corresponding to the image collected each time completely coincides with the historical image. This makes it difficult to achieve pixel-level alignment between images, hindering reliable comparison of disease features at the same position, and further affecting the accuracy of trend analysis. Currently, many schemes use GPS coordinates or IMU information for rough alignment, but still cannot meet the needs of high-precision disease time sequence monitoring.

[0005] Therefore, how to ensure image detection quality in complex lighting environments and achieve accurate alignment of cross-period images is a key problem to improve the accuracy and reliability of tunnel structure disease detection, which needs further research and solution. SUMMARY

[0006] The purpose of the embodiments of the present application is to provide a tunnel disease detection method and system based on a UAV, so as to at least solve the problems of unstable image quality under complex light environment conditions and difficulty in realizing accurate alignment and time sequence comparison of the same structure area in multiple flights in the existing tunnel disease detection process.

[0007] In order to achieve the above-mentioned purpose, the first aspect of the present application provides a tunnel disease detection method based on a UAV, the method comprising: controlling the UAV to fly along the tunnel to collect multi-channel image data of the structure surface, and adjusting the flight attitude based on the environmental light information to perform image light compensation; constructing a structure anchor point map based on historical collected images, combining the UAV flight pose information and the multi-channel image data after light compensation, and performing image alignment of each target anchor point; identifying the disease area in the aligned image, extracting the disease characteristics of each target anchor point this time, and updating the disease of each target anchor point this time into the time sequence characteristic data sequence of each target anchor point; wherein the time sequence characteristic data sequence of each target anchor point is the time sequence of the disease characteristics of each target anchor point in each detection round; based on the updated time sequence characteristic data sequence of each target anchor point, using Bayesian variable point detection to analyze the disease evolution trend of the tunnel.

[0008] Optionally, the control of the UAV to fly along the tunnel to collect multi-channel image data of the structure surface comprises: controlling the UAV to perform flight and image collection operations based on the collection route of the current tunnel, the collection route of the current tunnel being determined based on a pre-scanning operation performed in the current tunnel; the pre-scanning operation comprises collecting multi-channel image data of multiple positions in a flight state lower than a preset pre-scanning speed, and determining the brightness distribution, edge occlusion rate and texture complexity of each collection position based on each multi-channel image; wherein the brightness distribution is used to represent the illumination uniformity; the edge occlusion rate is used to represent the occlusion risk of the current field of view; the texture complexity includes texture entropy and edge density, and is used to represent the richness of structure details; the brightness distribution, edge occlusion rate and texture complexity of each collection position are mapped into a three-dimensional space heat map, and a target path scoring function is constructed according to the three-dimensional space heat map, so as to select the continuous path segment with the highest score as the image collection path of the current tunnel.

[0009] Optionally, the adjustment of the flight attitude based on the environmental light information to perform image light compensation comprises: performing image partition processing on the collected multi-channel image data, dividing the image into multiple pixel regions; calculating the brightness mean value and the gray variance of each pixel region respectively, and judging whether there is an exposure abnormal area according to a preset exposure threshold; when there is an exposure abnormal area, controlling the UAV to stay at the corresponding position and continuously collect multiple frames of multi-channel image data; performing geometric alignment processing on the multiple frames of image data, calculating the exposure weight based on the brightness difference, and generating enhanced multi-channel image data by pixel weighted fusion.

[0010] Optionally, the construction rule of the structure anchor atlas is: extracting a region with a texture change rate exceeding a preset change rate threshold from historical collected multi-channel image data as a candidate anchor region; generating a texture direction encoding graph, an edge gradient matrix and a frequency domain gray response graph for the candidate anchor region respectively; generating a texture fingerprint vector of a structure anchor based on the generated texture direction encoding graph, the edge gradient matrix and the frequency domain gray response graph, performing association of the texture fingerprint vector of the structure anchor and a flight attitude at a corresponding moment to construct a corresponding association relationship; and constructing the structure anchor atlas based on all association relationships.

[0011] Optionally, the image pair positioning of each target anchor point comprises: extracting a texture fingerprint vector from multi-channel image data corresponding to a current frame; performing multi-scale feature matching between the current texture fingerprint vector and a texture fingerprint vector of a structure anchor in the structure anchor atlas, and calculating a matching error; when the matching error exceeds a tolerance range, controlling the unmanned aerial vehicle to retreat to a previous attitude and re-collecting images; when the matching error is within the tolerance range, calculating an affine transformation matrix based on a corresponding relationship between the structure anchor in the structure anchor atlas and the target anchor point, and performing coordinate pair positioning processing on image data of the current frame based on the calculation result.

[0012] Optionally, the disease area is identified in the positioned image, and the disease feature of each target anchor point in this time is extracted, comprising: constructing a polar coordinate template region with the structure anchor corresponding to the target anchor point as the origin; performing edge extraction and texture direction clustering operations in the template region based on the multi-channel image data after light compensation to identify a crack area; calculating a crack average width of the crack, a texture direction entropy of the crack area and a definition index of the crack edge as crack feature data based on the identified crack area, and constructing a disease feature vector of the current detection round based on the crack feature data.

[0013] Optionally, the construction rule of the time sequence feature data sequence of each target anchor point is: the disease feature vectors extracted by each target anchor point in each round are composed into a time sequence feature data sequence in the order of collection time, the vectors comprising a crack average width vector, a texture direction entropy vector and an edge definition index vector; the time sequence feature data sequence further comprises an image collection time stamp and an image quality score of each round of detection.

[0014] Optionally, based on the updated time sequence characteristic data sequence of each target anchor point, a Bayesian change point detection is used to analyze the disease evolution trend of the tunnel, including: respectively extracting a crack average width sequence, a texture direction entropy sequence and an edge definition index sequence from the time sequence characteristic data sequence; using a Gaussian change model as a prior, a recursive Bayesian probability graph model is constructed; a one-dimensional time sequence model is constructed for the crack average width sequence, the texture direction entropy sequence and the edge definition index sequence, and a Gaussian variance change model is used as a prior to perform maximum a posteriori estimation to obtain a change point probability sequence corresponding to each data sequence; weighted fusion is performed on the change point probability sequence corresponding to each data sequence to obtain a disease evolution score of each target anchor point at each time point in the time sequence; based on the disease evolution score at each time point, disease evolution trend analysis is performed to obtain disease evolution analysis results of each target anchor point; and the disease evolution analysis results of each target anchor point are combined to obtain the disease evolution trend of the tunnel.

[0015] The second aspect of the present application provides a tunnel disease detection system based on a UAV, the system comprising: a collection unit configured to control the UAV to fly along the tunnel to collect multi-channel image data of the structure surface, and adjust the flight attitude based on ambient light information to perform image light compensation; a positioning unit configured to perform image positioning of each target anchor point based on a structure anchor point atlas constructed based on historical collected images, in combination with flight pose information of the UAV and the multi-channel image data after light compensation; a data updating unit configured to identify a disease area in the positioned image, extract disease characteristics of each target anchor point this time, and update the disease of each target anchor point this time into a time sequence characteristic data sequence corresponding to each target anchor point; wherein the time sequence characteristic data sequence of each target anchor point is a time sequence of disease characteristics of each target anchor point in each round of detection; and a trend evolution unit configured to analyze the disease evolution trend of the tunnel based on the updated time sequence characteristic data sequence of each target anchor point, using Bayesian change point detection.

[0016] In another aspect, the present application provides a computer readable storage medium having instructions stored thereon, which, when executed on a computer, cause the computer to perform the above-mentioned tunnel disease detection method based on a UAV.

[0017] By the technical scheme, the application introduces a multi-channel image acquisition and light perception adjustment mechanism, effectively alleviates the influence of complex light conditions of the tunnel on image quality, and improves the imaging clarity of the disease area. On this basis, a structure anchor point map is constructed combined with historical images, and flight pose information is fused to realize accurate alignment between images, so that the same structure position can be accurately calibrated and compared in multiple flights. Further, by organizing disease characteristics of each round according to the anchor points into time series data, and using a Bayesian variable point detection method to analyze the feature change trend, the quantitative identification and intelligent warning of the disease evolution process are realized. The alignment accuracy, time series comparability and risk judgment ability of the tunnel disease detection are improved as a whole.

[0018] Other features and advantages of the present application will be described in detail in the following detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0019] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and serve to explain the principles of the application. In the drawings:

[0020] Figure 1 is a step flow chart of a tunnel disease detection method based on a UAV provided by an embodiment of the application;

[0021] Figure 2 is a tunnel disease evolution score trend comparison chart provided by an embodiment of the application;

[0022] Figure 3 is a system structure chart of a tunnel disease detection system based on a UAV provided by an embodiment of the application. DETAILED DESCRIPTION

[0023] The specific embodiments of the application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to illustrate and explain the application, and are not used to limit the application.

[0024] Figure 1 is a method flow chart of a tunnel disease detection method based on a UAV provided by an embodiment of the application. As shown in Figure 1 , the application provides a tunnel disease detection method based on a UAV, which comprises the following steps:

[0025] Step S10: controlling the UAV to fly along the tunnel to collect multi-channel image data of the structure surface, and adjusting the flight attitude based on the ambient light information to perform image light compensation.

[0026] Specifically, the unmanned aerial vehicle is controlled to fly along the tunnel to collect multi-channel image data of the structure surface, including: controlling the unmanned aerial vehicle to perform flight and image collection operations based on a collection route of the current tunnel, the collection route of the current tunnel being determined based on a pre-scanning operation performed in the current tunnel; the pre-scanning operation includes collecting multi-channel image data of a plurality of positions in a flight state lower than a preset pre-scanning speed, and determining the brightness distribution, the edge occlusion rate and the texture complexity of each collection position based on each multi-channel image; wherein the brightness distribution is used to represent the illumination uniformity; the edge occlusion rate is used to represent the occlusion risk of the current field of view; the texture complexity includes texture entropy and edge density, and is used to represent the richness of structure details; the brightness distribution, the edge occlusion rate and the texture complexity of each collection position are mapped to a three-dimensional space heat map, and a target path scoring function is constructed according to the three-dimensional space heat map, so as to select the continuous path segment with the highest score as the image collection path of the current tunnel.

[0027] In the embodiment of the application, a pre-scanning operation is performed before formal image collection. The pre-scanning operation refers to controlling the unmanned aerial vehicle to preliminarily fly along the tunnel at a speed lower than a preset pre-scanning speed (for example, not more than 0.5 m / s), and collecting corresponding multi-channel image data at a plurality of spatial positions. The multi-channel image data usually includes a visible light channel, an infrared channel or other auxiliary imaging channels, which facilitates subsequent illumination compensation and structure identification. The position points collected by the pre-scanning flight can be uniformly distributed on the longitudinal path of the tunnel, with an interval of, for example, 1 meter or less, to ensure that the sampling resolution meets the downstream processing requirements.

[0028] For the image data obtained for each sampling point, quality evaluation of image illumination, structure details and occlusion conditions is performed respectively. Specifically, first, the image is converted to grayscale and analyzed for brightness distribution, and the brightness mean value and brightness standard deviation are calculated based on the pixel brightness values in each image. The smaller the brightness standard deviation, the more uniform the illumination, which is suitable for subsequent image recognition. The brightness standard deviation is defined as the illumination uniformity evaluation index.

[0029] Secondly, an image edge detection operator (such as Sobel or Canny) is used to extract the image edge profile, and the edge overlap ratio of the image peripheral region is counted to evaluate whether there is occlusion (such as pipelines, equipment and other structures). The occluded area often has a high edge coincidence degree and a chaotic gradient direction, and the higher the occlusion rate obtained by comprehensive evaluation, the lower the usability of the image at that position. The occlusion rate is used as the evaluation basis for the edge occlusion risk.

[0030] Then, based on texture analysis methods (such as Local Binary Pattern, LBP, or Gray Level Co-occurrence Matrix, GLCM), the texture features of the image are extracted, and the texture entropy index is calculated to represent the complexity of the structural details in the image. At the same time, the number of effective edge points per unit area of the image is counted as the edge density index. Both of them are used as the basis for measuring the "texture complexity" to quantify the image's ability to express structural topographic information.

[0031] The three indicators for each sampling position, i.e., brightness standard deviation (illumination uniformity), edge occlusion rate, and texture complexity (composed of texture entropy and edge density), are normalized and assigned weight coefficients, such as 0.4, 0.3, and 0.3, as inputs to the image quality scoring model, resulting in a comprehensive image quality score for each position.

[0032] The quality scores of each sampling position are interpolated or rasterized in a three-dimensional space (in the tunnel's internal coordinate system) to form a continuously distributed spatial image quality heat map. This heat map reflects the image acquisition availability at each position inside the entire tunnel and serves as the basis for subsequent path optimization.

[0033] Based on the image quality heat map, a target path scoring function is constructed. The scoring function can identify the highest-scoring continuous path segment in the heat map as the final image acquisition path based on path optimization strategies such as dynamic programming, A* search, or genetic algorithms. The preferred path segment should meet multiple conditions: the illumination uniformity index at most positions is below a set threshold (e.g., standard deviation < 15 gray levels), the occlusion rate is below a set upper limit (e.g., < 20%), the texture complexity is above the minimum recognition standard (e.g., entropy value > 1.5, edge density > 0.3), and it has good flight line feasibility (e.g., continuous obstacle-free flight distance is greater than 10 meters).

[0034] The selected image acquisition path is used as the flight trajectory for the formal sampling stage, and the UAV is controlled to fly strictly along this path and simultaneously complete multi-channel image acquisition. The above path selection mechanism not only effectively avoids lighting interference and occluded areas but also significantly improves the overall quality and structural feature expression ability of the collected images, laying a high-quality data foundation for subsequent structural disease identification and temporal comparison.

[0035] Preferably, the flight attitude is adjusted based on the ambient light information to perform image light compensation, including: performing image partitioning processing on the acquired multi-channel image data, dividing the image into a plurality of pixel regions; calculating the brightness mean and the gray variance of each pixel region respectively, and judging whether there is an exposure abnormal area according to a preset exposure threshold; when there is an exposure abnormal area, controlling the unmanned aerial vehicle to stop at the corresponding position and continuously acquire a plurality of frames of multi-channel image data; performing geometric alignment processing on the plurality of frames of image data, calculating the exposure weight based on the brightness difference, and generating enhanced multi-channel image data by pixel weighted fusion.

[0036] In the embodiment of the application, the multi-channel image data acquired by the unmanned aerial vehicle during flight is preprocessed to evaluate the image light distribution and find potential exposure abnormal areas. Specifically, each frame of image is divided into a plurality of pixel regions, for example, using a fixed size grid partitioning method, the image is divided into a plurality of sub-regions with a size of 32x32 or 64x64 pixels, ensuring that each region contains enough pixels for brightness statistical analysis.

[0037] The brightness mean (such as based on Y channel or grayscale image) and the gray variance of each pixel region are calculated respectively, the brightness mean is used to judge the overall brightness level of the region, and the gray variance is used to evaluate the local light contrast of the region. For the brightness mean of each region, compare it with the set exposure judgment threshold (for example, the overexposure threshold is 240 and the underexposure threshold is 30), if the brightness mean of a region exceeds the overexposure threshold or is lower than the underexposure threshold, it is marked as an exposure abnormal area. By traversing each region, the exposure abnormal area distribution map of the current image can be obtained.

[0038] If there is an exposure abnormal area in a frame of image, and the position of the area is located in a structure detection key area (such as near the structure anchor point or the texture complex section), it is preferred to keep the flight stable state at this position, and continuously acquire a plurality of frames of multi-channel image data for light enhancement. The number of frames can be determined according to the preset sampling number (for example, 3-5 frames) or until the brightness distribution converges.

[0039] For the acquired multiple frames of images, geometric alignment processing needs to be performed first to ensure that the same structure region has consistent position in different images. The alignment processing can use feature point based image registration algorithm, such as SIFT, ORB and other methods to extract key points, and then use RANSAC to calculate homography matrix to realize image registration, or use dense optical flow method to fine-tune in weak texture area to improve the alignment accuracy. After registration, all frames of images are mapped to a common reference coordinate.

[0040] Further, based on the brightness value of each pixel in each frame of image, the exposure weight of the corresponding pixel in the fusion process is calculated. The calculation of the exposure weight can adopt a bilateral weighting strategy: pixels with brightness near the middle value (such as brightness of 120-180) are given a higher weight, and pixels with brightness in the high or low area are given a lower weight, so as to avoid the interference of overexposure or underexposure information on the fusion result. The exposure weight can be expressed as:

[0041] ;

[0042] wherein, is the pixel brightness, is the target exposure level (such as 150), is the brightness tolerance parameter (such as 30). After the exposure weight of all pixel positions is calculated, the pixel-level weighted fusion is performed on all frame images, and the fusion method can be a weighted average operation performed on each channel respectively. The value of each pixel in the final output image is the weighted average of the pixel values in each frame, generating a multi-channel image with balanced brightness and enhanced details, which is the enhanced image result at this position.

[0043] Step S20: based on the structure anchor atlas constructed based on the historical acquisition images, combined with the flight pose information of the unmanned aerial vehicle and the multi-channel image data after light compensation, the image alignment of each target anchor is performed.

[0044] Specifically, the construction rule of the structure anchor atlas is: extracting a region with a texture change rate exceeding a preset change rate threshold from the multi-channel image data obtained from historical acquisition as a candidate anchor region; generating a texture direction encoding graph, an edge gradient matrix and a frequency domain gray response graph for the candidate anchor region respectively; based on the generated texture direction encoding graph, edge gradient matrix and frequency domain gray response graph, a texture fingerprint vector of a structure anchor is formed, the texture fingerprint vector of the structure anchor is associated with the flight attitude at the corresponding time to construct a corresponding association relationship; and the structure anchor atlas is constructed based on all association relationships.

[0045] Further, the image alignment of each target anchor includes: extracting a texture fingerprint vector from the multi-channel image data corresponding to the current frame; performing multi-scale feature matching between the current texture fingerprint vector and the texture fingerprint vector of the structure anchor in the structure anchor atlas, and calculating a matching error; when the matching error exceeds a tolerance range, controlling the unmanned aerial vehicle to retreat to the last pose and reacquire images; when the matching error is within the tolerance range, calculating an affine transformation matrix based on the corresponding relationship between the structure anchor in the structure anchor atlas and the target anchor, and performing coordinate alignment processing on the image data of the current frame based on the calculation result.

[0046] In the embodiment of the present application, the construction process of the structure anchor atlas includes two key stages: anchor region screening and anchor feature extraction and fingerprint encoding.

[0047] In the anchor point area screening stage, from the multi-channel image data that has been collected in the historical detection process, a region with high structural feature representation and stable identifiability is identified and selected as a candidate anchor point. Considering that the internal structure of the tunnel is mostly a repetitive, high-texture curved surface or concrete lining surface, a region randomly sampled in the image may lack sufficient distinguishability and be difficult to track stably for a long time. Therefore, the texture change rate needs to be introduced as a preliminary screening index.

[0048] The calculation of the texture change rate can use a sliding window to traverse the image, and the window size is taken as 64x64 pixels, for example. The texture direction gradient change value of the image inside the window is calculated. The specific method is as follows: the Histogram of Oriented Gradient (HOG) or structure tensor feature is used to statistically calculate the main direction distribution and variance change rate of the local region of the image. Then, the same region image collected in the historical rounds of inspection of the tunnel section is compared, the texture direction difference variance in each round of image is calculated, and the difference variance is taken as the texture change rate of the region. If the texture change rate of a region is stable for a long time, the direction distribution is clear, and the contrast in the image is sufficient, the region can be identified as a candidate structural anchor point.

[0049] After obtaining the candidate anchor point region, feature extraction is continued to construct a texture fingerprint vector. Specifically, the following three types of image features representing structural characteristics are generated for each candidate region respectively:

[0050] 1) Texture direction encoding map: local binary pattern (LBP) or direction gradient encoding (such as Gabor filtering) is used to process the region image to extract its local texture direction features. The encoded result is a fixed-length direction feature vector with certain rotation robustness;

[0051] 2) Edge gradient matrix: the edge intensity and direction information of the image is obtained using gradient operators such as Sobel, Scharr or Laplacian, and the gradient mean, maximum and standard deviation of the region in the horizontal and vertical directions are calculated to form a numerical matrix reflecting the boundary shape;

[0052] 3) Frequency domain gray response map: fast Fourier transform (FFT) is performed on the candidate region to calculate the response strength and distribution of the frequency spectrum in the medium and high frequency bands, and a frequency domain feature vector reflecting the image details and texture periodicity is obtained.

[0053] The three types of features are spliced and normalized to form a texture fingerprint vector of the anchor point. To ensure tracking accuracy, the texture fingerprint vector is bound to the flight attitude data (including position coordinates XYZ, pitch angle Pitch, yaw angle Yaw, and roll angle Roll) of the corresponding image acquisition time to form a one-to-one spatial correlation. The texture fingerprint vectors of all anchor points and their associated flight poses form a structural anchor point map, which constitutes a reference database for subsequent image registration.

[0054] In subsequent inspection flights, the current image and the anchor point map need to be registered. Specifically, from the multi-channel image data after illumination compensation of the current frame, the image region is extracted according to the same partitioning rule as the anchor point map, and the texture fingerprint vector of the target region is generated. The fingerprint generation method is consistent with the historical map to ensure that the vector structure is completely comparable.

[0055] After extraction, a multi-scale feature matching strategy is used to find the closest historical anchor point in the structural anchor point map to the current frame texture fingerprint vector. The matching can use vector Euclidean distance, cosine similarity, or KL divergence as the measurement function to obtain the matching error value of each matching candidate.

[0056] To improve robustness, the matching process can be repeated at multiple scales (such as 64x64, 96x96, 128x128 windows), and the scale with the smallest matching error is finally used to obtain the best correspondence between the target anchor point in the current image and the historical anchor point in the structural anchor point map. When the matching error exceeds the preset tolerance (such as a similarity threshold of 0.3), it means that the image region may be affected by occlusion, blur, or flight error. At this time, the image re-sampling logic should be triggered to control the UAV to back up to the previous pose point according to the original trajectory to reacquire the image to ensure that the anchor point region image quality meets the standard.

[0057] If the matching error is within the tolerance range, the current image anchor recognition is considered successful. At this time, combined with the spatial mapping relationship between the anchor point in the structural anchor point map and its corresponding target anchor point (which can be derived through flight attitude recording and acquisition coordinates), the affine transformation matrix between the current frame image and the target reference image is calculated. Affine transformation calculation can be based on matching point pair coordinates fitting, such as using least squares method or RANSAC optimization to obtain the transformation matrix. Based on the transformation matrix, the spatial coordinate registration of the current frame image data is performed to strictly align the target anchor point position of the current image to the historical reference frame coordinate system.

[0058] The alignment mechanism avoids the computational burden of forced registration of the whole image, and significantly improves the alignment accuracy and efficiency by using the texture stability of local anchor points. Especially in the analysis of disease evolution trend of tunnel structure, through accurate image alignment, it can ensure that the disease features in images of different periods are extracted in a unified coordinate, greatly improving the accuracy and reliability of the analysis results of disease growth and crack expansion.

[0059] Step S30: identifying the disease area in the aligned image, extracting the disease features of each target anchor point this time, and updating the disease of each target anchor point this time into the time sequence feature data sequence of each target anchor point.

[0060] Specifically, the time sequence feature data sequence of each target anchor point is the time sequence of the disease features of each target anchor point in each round of detection.

[0061] Further, identifying the disease area in the aligned image and extracting the disease features of each target anchor point this time includes: constructing a polar coordinate template area with the structure anchor point corresponding to the target anchor point as the origin; performing edge extraction and texture direction clustering operations in the template area based on the multi-channel image data after illumination compensation to identify the crack area; calculating the crack average width of the crack, the texture direction entropy of the crack area, and the clarity index of the crack edge as the crack feature data based on the identified crack area, and constructing the disease feature vector of the current detection round based on the crack feature data.

[0062] Further, the construction rule of the time sequence feature data sequence of each target anchor point is to arrange the disease feature vectors of each target anchor point extracted in each round in time sequence to form a time sequence feature data sequence, the vector including a crack average width vector, a texture direction entropy vector, and an edge clarity index vector; the time sequence feature data sequence also includes an image acquisition time stamp and an image quality score of each round of detection.

[0063] In the embodiment of the present application, considering that there are factors such as curved surface deformation, angle difference, and texture disturbance in actual tunnel structures, directly using a regular rectangular window may introduce non-structural noise. Therefore, when identifying the disease, a local polar coordinate template area is constructed with the center of the structure anchor point corresponding to the target anchor point as the origin in the image coordinate system. The polar coordinate template can be set as an anchor point with a radius r of 50 to 100 pixels (flexibly set according to the image resolution), an angle range covering 0° to 360°, and divided into several angle sectors (such as one sector every 30°), forming an identification window with direction perception ability.

[0064] Within this template region, a crack candidate region identification operation is performed. This process is based on the post-illumination compensation multi-channel image data, first performing a standard edge enhancement and extraction operation on the image, such as applying a Canny edge detection algorithm, in conjunction with a high-pass filter to enhance texture boundaries. To eliminate non-structural redundant edge interference, after edge detection, a direction clustering algorithm (such as based on k-means or DBSCAN) is used to cluster the extracted edge segments by the main direction, obtaining a set of line segments with continuous direction. If the direction concentration of a group of edge segments after clustering is high, the coverage area is narrow and the gray scale contrast is strong, it can be preliminarily determined as a suspected crack region.

[0065] For each identified crack region, its key structural features are extracted to construct the disease feature data of this round of detection. The crack features include the following three:

[0066] 1) Crack average width: a vertical sampling line is established along the main direction of the crack, the pixel distance between the crack boundaries on this line is calculated, and the average value is obtained by averaging multiple sampling positions, to obtain the average width of the crack, in units of pixel number (or converted to millimeters, according to the flight height and camera resolution).

[0067] 2) Texture direction entropy: a texture window is established in the vicinity of the crack, and local direction information is counted using a Gabor filter or a histogram of oriented gradients (HOG), and the entropy value of the direction distribution is calculated. The greater the direction entropy, the more complex the direction structure of the region, which may correspond to multiple crack intersections, denudation and other multi-morphology damage.

[0068] 3) Edge sharpness index: obtained by calculating the average value and standard deviation of the gray scale gradient of the crack boundary region. The specific method is as follows: extract a 5-pixel-wide image band on both sides of the crack boundary, and calculate the average value and variance of the gray scale gradient respectively. If the edge gradient is large and the variance is low, it indicates that the edge is clear and the imaging quality is high, with high detection reliability.

[0069] The above three indicators are combined into a disease feature vector, denoted as:

[0070] ;

[0071] wherein, is the crack average width, is the texture direction entropy, is the edge sharpness.

[0072] The vector is the disease structure description data of the i-th target anchor point in the current detection round. To support long-term tracking and trend analysis, a continuous time series feature data sequence needs to be constructed for each target anchor point. The construction rule is to record the disease feature vector extracted in each round of inspection for each target anchor point, and organize it into a time series list in chronological order. Each round of disease feature includes: crack average width vector (unit: mm or pixel), texture direction entropy vector (dimensionless), edge sharpness index vector (unit: gray gradient intensity), image acquisition timestamp (format: YYYY-MM-DD hh:mm:ss), and image quality score (which can be calculated based on sharpness, exposure, and occlusion rate, 0-1 score).

[0073] All round feature vectors are arranged in ascending order of time to form the complete time series sequence of the anchor point. This structure not only guarantees the time continuity of the data, but also lays a structured data foundation for subsequent statistical modeling of change point detection and trend judgment. It should be noted that, in order to ensure the stability of sequence analysis, it is preferred to establish a trend analysis model on samples with image scores higher than a set threshold (such as q>0.7). If the quality score of a certain frame of image is low, it can be removed or given a lower weight in the post-processing stage to reduce its impact on the accuracy of disease evolution trend identification.

[0074] Step S40: Based on the updated time series feature data sequence of each target anchor point, the Bayesian change point detection is used to analyze the disease evolution trend of the tunnel.

[0075] Specifically, the crack average width sequence, the texture direction entropy sequence, and the edge sharpness index sequence are extracted from the time series feature data sequence respectively; a recursive Bayesian probability graph model is constructed using a Gaussian change model as a prior; one-dimensional time series models are constructed for the crack average width sequence, the texture direction entropy sequence, and the edge sharpness index sequence, and a Gaussian variance change model is used as a prior to perform maximum a posteriori estimation to obtain the change point probability sequence corresponding to each data sequence; weighted fusion is performed on the change point probability sequence corresponding to each data sequence to obtain the disease evolution score of the target anchor point at each time point in the time sequence; based on the disease evolution score at each time point, disease evolution trend analysis is performed to obtain the disease evolution analysis result of each target anchor point; and the disease evolution analysis results of each target anchor point are combined to obtain the disease evolution trend of the tunnel.

[0076] In the embodiments of the present application, time series data corresponding to three types of key physical features are extracted from the complete time sequence feature data sequence of each target anchor point, including: a crack average width sequence (indicating the spatial expansion trend of the disease), a texture direction entropy sequence (indicating the disease complexity or evolution activity), and an edge sharpness index sequence (indicating the time variation of the image imaging quality or edge differentiation degree). The above three types of sequences are arranged in ascending order of image collection time stamp to form a one-dimensional time sequence signal.

[0077] Secondly, a Bayesian change point detection is used as the core analysis method to construct a time sequence model for each of the above three sequences. Considering that there is often a non-stationary process in the development process of the tunnel structure disease (for example, the crack slowly expands at a certain period and then rapidly increases), a Gaussian distribution change model is used as a prior assumption for each feature sequence, that is, in a statistical sense, the disease characteristics obey a certain fixed Gaussian distribution in a stable interval, but the mean or variance of the distribution will jump significantly at the change point.

[0078] In order to establish a recursive calculation structure, a Bayesian online change point detection (BOCPD) is used as a change point analysis framework. This method allows real-time evaluation of whether there is a distribution jump at each newly added data point. The specific process is as follows:

[0079] 1) For any feature sequence , initialize the running length distribution , and set the initial value as .

[0080] 2) Set a potential change point position for each time point t, that is, the time length since the last change point.

[0081] 3) For each time point, recursively update , and use a Gaussian variance change model as the prior of the observation distribution.

[0082] 4) In the Bayesian update, the posterior estimation is performed by using the running length distribution at the previous time and the observation likelihood function .

[0083] 5) The probability of the existence of a change point at each time point is obtained by maximum a posteriori (MAP) estimation .

[0084] The above calculation is applied to the crack width sequence W(t), the texture direction entropy sequence E(t), and the edge sharpness index sequence C(t) to obtain three independent change point probability sequences:

[0085] 1) : Change point probability corresponding to crack width change.

[0086] 2) : Change point probability corresponding to texture direction entropy change.

[0087] 3) : Change point probability corresponding to edge sharpness change.

[0088] To improve the robustness of the overall judgment, the above three change point probability sequences need to be weighted and fused to obtain an evolution score supported by multiple feature dimensions. The specific fusion method is as follows:

[0089] 2) Normalize the three change point probability sequences to make their numerical values uniformly distributed in the [0, 1] interval.

[0090] 3) Set feature weight coefficients , satisfying , for example, set to 0.4, 0.3, 0.3 according to feature confidence or data quality evaluation.

[0091] 4) Perform weighted fusion calculation: ;

[0092] wherein is the disease evolution score of the target anchor point at time point t.

[0093] The above fusion score reflects whether the disease features simultaneously present a jump trend in multiple dimensions. Analyze the continuous high score segment in the time series to determine whether there is a stable evolution trend. To further quantify the evolution trend, set a score threshold , for example , when consecutive time points (e.g. 3) satisfy , it can be determined that the target anchor point enters a significant evolution state. Further extract the feature increment indicators corresponding to these time periods to calculate crack growth rate, direction entropy growth rate, etc. to supplement the quantitative description.

[0094] After completing the disease evolution score of each target anchor point, aggregate the trend determination results of all target anchor points within the current detection period to perform a disease trend synthesis analysis on the whole tunnel scale. Construct an anchor point distribution map according to the spatial distribution of the anchor points in the tunnel; mark the disease trend type of each target anchor point (such as: stable, growth, fluctuation); generate an evolution risk heat map based on the evolution score heat;

[0095] Statistical evolution anchor point proportion, spatial aggregation, etc. to determine whether there is a disease concentrated development area;

[0096] Output the structure paragraph level trend report, including "review area" "stable area" "suspected acceleration area" and other classification suggestions.

[0097] Embodiments:

[0098] Select a certain city subway section as the detection object, and perform 20 rounds of fixed-point image acquisition on a target structure anchor point with known early fine cracks. In each round of detection, the unmanned aerial vehicle platform collects multi-channel image data of the target anchor point according to the predetermined trajectory, and extracts three key feature indicators of the anchor point region after image alignment, which are: crack average width (unit: millimeter); crack region texture direction entropy (measures structure complexity); crack edge definition index (measures boundary imaging quality).

[0099] The above three indicators form a three-dimensional disease feature vector, which is written into the time sequence feature data sequence of the corresponding round, and is attached with an image acquisition timestamp and a quality score value, ensuring sequence continuity and data availability.

[0100] In the scheme of the present application, independent one-dimensional feature sequences are constructed for crack width, texture direction entropy and edge definition, a Gaussian change model is used as a prior hypothesis, and a Bayesian online change point detection algorithm (BOCPD) is introduced for probability inference. Based on the historical distribution of the target feature, the algorithm determines whether there is a statistically significant distribution change at the current time, so as to identify potential "disease jump points". Each one-dimensional feature sequence generates an independent change point probability curve, and a weighted fusion calculation of disease evolution score is performed on this basis. The weighting coefficients can be set as width: entropy: definition = 0.4:0.3:0.3, which can be adjusted according to application requirements.

[0101] As shown in Figure 2 , between the 8th and 12th detection rounds, the score of the present application rises rapidly and breaks through the threshold at the 10th round, successfully capturing the state of joint mutation of crack features; in contrast, the score of the comparative scheme rises slowly, and gradually approaches or is slightly higher than 0.7 after the 14th round, failing to timely reflect the real disease evolution trend; the score change of the present scheme is more steep and sensitive to fluctuations, which can clearly locate the change point period, while the comparative scheme presents a fuzzy transition in change point identification, lacking judgment basis. Combined with the on-site manual review results, visible width growth and morphological complexity characteristics of the cracks in this area are indeed observed after the 10th round, further verifying the forward-looking nature of the method in trend identification. In addition, the score of the present scheme is relatively stable in the early detection rounds (1-7 rounds), indicating that the method has the ability to identify non-evolution stages stably, avoiding false positives or hypersensitive responses, thereby having good practicality and reliability in actual engineering.

[0102] Figure 3is a system structure diagram of a tunnel disease detection system based on a UAV provided by an embodiment of the present application. As shown in Figure 3 An embodiment of the present application provides a tunnel disease detection system based on a UAV, which comprises: a collection unit configured to control a UAV to collect multi-channel image data of a structure surface while flying along a tunnel, and adjust a flight attitude based on ambient light information to perform image light compensation; a positioning unit configured to perform image positioning of each target anchor point based on a structure anchor point map constructed based on historical collected images, in combination with flight pose information of the UAV and the multi-channel image data after light compensation; a data updating unit configured to identify a disease area in the positioned image, extract disease features of each target anchor point in this round, and update the disease of each target anchor point in this round into a time sequence feature data sequence of each target anchor point; and a trend evolution unit configured to analyze a disease evolution trend of the tunnel based on the updated time sequence feature data sequence of each target anchor point by using Bayesian variable point detection.

[0103] An embodiment of the present application further provides a computer readable storage medium, which stores instructions, and the instructions make a computer execute the tunnel disease detection method based on a UAV when the computer runs.

[0104] Those skilled in the art can understand that all or part of the steps of the method for implementing the above-mentioned embodiments can be completed by programs instructing related hardware, the programs are stored in a storage medium, and the programs include a plurality of instructions for making a single-chip microcomputer, a chip or a processor execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk and various storage medium capable of storing program codes.

[0105] The optional embodiments of the present application are described in detail above in combination with the drawings, but the embodiments of the present application are not limited to the specific details in the above-mentioned embodiments, and various simple modifications can be made to the technical solutions of the embodiments of the present application within the technical concept of the embodiments of the present application, and these simple modifications all belong to the protection scope of the embodiments of the present application. In addition, it should be noted that each specific technical feature described in the above-mentioned specific embodiments can be combined in any appropriate manner without contradiction. In order to avoid unnecessary repetition, the embodiments of the present application will not further describe various possible combination manners.

[0106] Besides, the various embodiments of the present application can be combined arbitrarily, as long as it does not violate the idea of the embodiments of the present application, which should be considered as the disclosed content of the embodiments of the present application.

Claims

1. A method for detecting tunnel disease based on a UAV, characterized in that, The method comprises: controlling the unmanned aerial vehicle to collect multi-channel image data of the structure surface along the tunnel, and adjusting the flight attitude based on the ambient light information to perform image light compensation; based on the structure anchor point map constructed based on historical collected images, combined with the flight pose information of the unmanned aerial vehicle and the multi-channel image data after light compensation, performing image alignment of each target anchor point; identify the disease area in the aligned image, extract the disease characteristics of each target anchor point this time, and update the disease of each target anchor point this time into the time sequence characteristic data sequence of each target anchor point; wherein, the time sequence characteristic data sequence of each target anchor point is the time sequence of the disease characteristics of each target anchor point in each round of detection; based on the updated time sequence characteristic data sequence of each target anchor point, use Bayesian variable point detection to analyze the disease evolution trend of the tunnel.

2. The method of claim 1, wherein, controlling the unmanned aerial vehicle to collect multi-channel image data of the structure surface along the tunnel, comprising: controlling the unmanned aerial vehicle to perform flight and image collection operations based on the collection route of the current tunnel, wherein the collection route of the current tunnel is determined based on the pre-scanning operation performed in the current tunnel; the pre-scanning operation comprises collecting multi-channel image data of multiple positions in a flight state lower than a preset pre-scanning speed, and determining the brightness distribution, edge occlusion rate and texture complexity of each collection position based on each multi-channel image; wherein, the brightness distribution is used to represent the illumination uniformity; the edge occlusion rate is used to represent the occlusion risk of the current field of view; the texture complexity includes texture entropy and edge density, and is used to represent the richness of structure details; map the brightness distribution, edge occlusion rate and texture complexity of each collection position to a three-dimensional space heat map, and construct a target path scoring function according to the three-dimensional space heat map to select the continuous path segment with the highest score as the image collection path of the current tunnel.

3. The method of claim 1, wherein, adjusting the flight attitude based on the ambient light information to perform image light compensation, comprising: performing image partition processing on the collected multi-channel image data, dividing the image into multiple pixel regions; calculate the brightness mean and gray variance of each pixel region respectively, and determine whether there is an exposure abnormal area according to the preset exposure threshold; when there is an exposure abnormal area, control the unmanned aerial vehicle to stop at the corresponding position and continuously collect multiple frames of multi-channel image data; performing geometric alignment processing on the multi-frame image data, calculating the exposure weight based on the brightness difference, and generating enhanced multi-channel image data by pixel weighted fusion.

4. The method of claim 1, wherein, The construction rule of the structure anchor point map is: extract the region with a texture change rate exceeding a preset change rate threshold from the historical collected multi-channel image data as a candidate anchor point region; generate a texture direction encoding map, an edge gradient matrix and a frequency domain gray response map for the candidate anchor point region respectively; based on the generated texture direction encoding map, edge gradient matrix and frequency domain gray response map, a texture fingerprint vector of a structure anchor point is formed, and the texture fingerprint vector of the structure anchor point is associated with the flight attitude at the corresponding time to construct the corresponding association relationship; construct the structure anchor point map based on all association relationships.

5. The method of claim 4, wherein, performing image alignment of each target anchor point, comprising: extracting a texture fingerprint vector from multi-channel image data corresponding to a current frame; performing multi-scale feature matching between the current texture fingerprint vector and texture fingerprint vectors of structural anchor points in a structural anchor point atlas, and calculating a matching error; when the matching error exceeds a tolerance range, controlling the UAV to retreat to a previous pose and reacquire image data; when the matching error is within the tolerance range, calculating an affine transformation matrix based on a correspondence relationship between the structural anchor points in the structural anchor point atlas and target anchor points, and performing coordinate alignment processing on image data of the current frame based on the calculation result.

6. The method of claim 4, wherein, identifying a disease area in the aligned image and extracting disease features of each target anchor point in this round, including: constructing a polar coordinate template region with a structural anchor point corresponding to the target anchor point as the origin; performing edge extraction and texture direction clustering operations in the polar coordinate template region based on the multi-channel image data after light compensation, and identifying a crack region; based on the identified crack region, calculating a crack average width, a texture direction entropy of the crack region, and a definition index of the crack edge as crack feature data, and forming a disease feature vector of the current detection round based on the crack feature data.

7. The method of claim 6, wherein, The construction rule of the time sequence feature data sequence of each target anchor point is: the disease feature vectors extracted by each target anchor point in each round are arranged in time sequence to form a time sequence feature data sequence, and the disease feature vectors include a crack average width vector, a texture direction entropy vector, and an edge definition index vector; the time sequence feature data sequence also includes an image acquisition timestamp and an image quality score of each round of detection.

8. The method of claim 7, wherein, Based on the updated time sequence feature data sequence of each target anchor point, the Bayesian change point detection is used to analyze the disease evolution trend of the tunnel, including: extracting a crack average width sequence, a texture direction entropy sequence, and an edge definition index sequence from the time sequence feature data sequence respectively; using a Gaussian change model as a prior, a recursive Bayesian probability graph model is constructed; one-dimensional time series models are constructed for the crack average width sequence, the texture direction entropy sequence, and the edge definition index sequence respectively, and a Gaussian variance change model is used as a prior to perform maximum a posteriori estimation, obtaining a change point probability sequence corresponding to each data sequence; performing weighted fusion on the change point probability sequences corresponding to each data sequence to obtain a disease evolution score of each target anchor point at each time point in the time sequence; based on the disease evolution scores at each time point, disease evolution trend analysis is performed to obtain disease evolution analysis results of each target anchor point; combining the disease evolution analysis results of each target anchor point to obtain the disease evolution trend of the tunnel. 9.A UAV-based tunnel disease detection system, characterized in that, The system includes: an acquisition unit configured to control a UAV to fly along a tunnel to acquire multi-channel image data of a structural surface, and adjust a flight attitude based on ambient light information to perform image light compensation; an alignment unit configured to perform image alignment of each target anchor point based on a structural anchor point atlas constructed based on historical acquired images, in combination with flight pose information of the UAV and the multi-channel image data after light compensation; The data updating unit is configured to identify a disease area in the image of the site, extract disease features of each target anchor point in the current round, and update the disease of each target anchor point in the current round into a time sequence feature data sequence corresponding to each target anchor point. The time sequence feature data sequence of each target anchor point is a time sequence of disease features of each target anchor point in each round of detection. The trend evolution unit is configured to analyze the disease evolution trend of the tunnel by using Bayesian variable point detection based on the updated time sequence feature data sequence of each target anchor point.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores instructions, which, when executed on a computer, cause the computer to perform the unmanned aerial vehicle based tunnel disease detection method of any one of claims 1-8.

Images