Search

CN-122016824-A - Tunnel surface crack real-time detection method based on image recognition

CN122016824ACN 122016824 ACN122016824 ACN 122016824ACN-122016824-A

Abstract

The invention belongs to the technical field of image data processing and structure detection, and particularly relates to a tunnel surface crack real-time detection method based on image recognition. The method comprises the steps of collecting tunnel surface images and motion information through multi-sensor fusion, estimating pose parameters in real time by utilizing a visual synchronous positioning and map construction algorithm, constructing a reverse motion compensation model according to the pose parameters to eliminate motion blur and perspective distortion generated by shake of a detection platform, generating a high-definition standardized image, automatically extracting crack geometric outlines and pixel mask features by utilizing a depth semantic segmentation network, projecting crack coordinates to a global coordinate system, and realizing incremental comparison of crack degradation trends by means of spatial alignment with historical inspection data. The invention solves the problems of image blurring and low positioning precision in a dynamic environment, realizes sub-centimeter global positioning and cross-period millimeter level change monitoring, provides reliable data support for long-term security assessment of a tunnel structure, and improves the automation degree and decision efficiency of inspection operation.

Inventors

  • He Huguo
  • QIN GUANGRONG
  • LI KAI
  • BAI JIANWEI
  • Dai Dongyu
  • LIU FENGXIANG
  • WANG WENYUE
  • ZHANG LIQIANG

Assignees

  • 中铁一局集团铁路建设有限公司
  • 中铁一局集团有限公司

Dates

Publication Date
20260512
Application Date
20260413

Claims (10)

  1. 1. The tunnel surface crack real-time detection method based on image recognition is characterized by comprising the following steps of: Step 1, constructing a multi-sensor fusion dynamic sensing system, and synchronously triggering a data acquisition instruction in the moving process of a detection platform by utilizing a high-resolution imaging unit, an inertia measurement unit and a laser ranging unit which are arranged on the detection platform to acquire continuous original sequence images, instantaneous angular velocity information, instantaneous acceleration information and preset distance data of the detection platform to the inner wall of the tunnel; Step 2, executing a front-end mileage calculation method based on visual synchronous positioning and map construction, and estimating six-degree-of-freedom pose parameters of a detection platform in a tunnel local space coordinate system in real time by extracting characteristic points in an original sequence image and performing inter-frame matching and combining motion increment output by an inertia measurement unit, wherein the six-degree-of-freedom pose parameters comprise a three-dimensional rotation matrix and a three-dimensional translation vector; Step 3, implementing dynamic drift compensation and image de-distortion treatment, calculating the tiny vibration displacement of the detection platform at the exposure moment by using the pose parameters estimated in the step 2, constructing a reverse motion compensation model, and performing pixel-level resampling and geometric correction on the original sequence image to generate a high-definition standardized tunnel surface image; Step 4, executing depth semantic segmentation and feature extraction logic, inputting a standardized tunnel surface image into a preset crack detection neural network model, automatically identifying and extracting geometric outline features, width features and length features of cracks, and generating a crack mask image with pixel-level labels; And 5, realizing sub-centimeter global positioning and historical evolution monitoring, mapping pixel coordinates in a crack mask image into a global world coordinate system through projective transformation by using a tunnel sparse point cloud map generated by synchronous positioning and map construction, calculating absolute physical coordinates of the crack, performing space alignment and increment comparison on a current detection result and historical inspection data stored in a preset database, and judging the degradation trend of the crack.
  2. 2. The method for detecting the surface cracks of the tunnel based on image recognition according to claim 1, wherein in the step 1, the high-resolution imaging unit comprises a plurality of groups of industrial cameras which are arranged in a fan shape along the circumferential direction of the detection platform, and a predetermined field of view overlapping area is kept between every two adjacent groups of cameras, wherein the width of the field of view overlapping area is 10% to 20% of the width of a single image; the triggering frequency of the high-resolution imaging unit and the running speed of the detection platform form closed-loop feedback regulation, the running speed is monitored in real time through a high-line number incremental encoder arranged on the running wheel of the detection platform, a speed signal is fed back to a central controller, and when the running speed is increased, the central controller automatically increases the triggering frequency so as to keep the displacement increment of the adjacent images in a longitudinal space constant; the sampling frequency of the inertial measurement unit is set to be more than two hundred hertz, and the micro angular velocity disturbance and the instantaneous linear acceleration change of the detection platform in the three-axis direction are recorded; The method further comprises the compensation treatment of ambient light, wherein in the image acquisition process, a plurality of groups of high-power constant-current illumination lamps are synchronously started, and diffusion sheets are arranged at the front ends of the plurality of groups of high-power constant-current illumination lamps so that output light rays are uniformly distributed on the wall surface of the tunnel; The multi-sensor fusion dynamic sensing system also utilizes a digital temperature and humidity sensor integrated on the detection platform to synchronously monitor environmental temperature and humidity parameters, and the current environmental temperature value and relative humidity value are stored into corresponding image information files as metadata additional attributes while crack images are acquired.
  3. 3. The method for detecting the surface cracks of the tunnel based on the image recognition according to claim 2, wherein in the step 2, when the front-end odometer constructed by the visual synchronous positioning and the map processes the image, the color image is firstly converted into a single-channel gray-scale image, and the finite-contrast self-adaptive histogram equalization processing is performed; Feature point extraction adopts feature descriptors with rotation invariance and scale invariance, and extremum points are searched in parallel on a plurality of layers from original resolution to multi-level downsampling resolution by establishing a feature point pyramid structure; The inter-frame matching process adopts a bidirectional optical flow tracking technology and combines a random sampling consistency algorithm to calculate the pixel motion vector of the current frame relative to the previous frame, reversely calculates the displacement of the previous frame relative to the current frame, reserves characteristic points when the bidirectional error is smaller than a preset threshold value, utilizes iteration to select a minimum sample set to construct a homography matrix, and eliminates outliers which do not accord with geometric constraint; The six-degree-of-freedom pose parameter estimation process further comprises a back-end optimization link, a factor graph model is constructed, a visual re-projection error, an inertial measurement unit pre-integration error and a transverse distance constraint item provided by a laser ranging unit are used as optimization objective functions, a Levenberg Marquardt algorithm is utilized to carry out iterative smoothing processing on a motion track of a detection platform in a section of sliding window, the sum of squares of residual errors between all observed data and the motion model is minimized, and accumulated drift errors generated by a sensor are eliminated.
  4. 4. The method for detecting tunnel surface cracks based on image recognition according to claim 3, wherein in the step 2, the inertia measurement unit pre-integration technology performs continuous time integration operation on high-frequency acceleration data and angular velocity data in a time span between two adjacent visual image frames to obtain a relative velocity increment, a displacement increment and an angle conversion amount of the detection platform in the time period, and counteracts the influence of instantaneous impact force generated by a vehicle at a track irregularity on a pose track; in the pose estimation process, the method also introduces an edge constraint term, extracts longitudinal and circumferential joint lines between lining blocks by using a Canni edge detection operator, establishes line feature descriptors and performs line matching, and adds a reprojection error of the line features into a factor graph optimization model; For the rolling shutter effect of the camera, the rotation angle increment of the camera in each line of scanning time is calculated by utilizing the high-frequency angular velocity data provided by the inertial measurement unit, the progressive scanning line compensation is carried out on the image, and the deformation generated by the rolling shutter is corrected.
  5. 5. The method for detecting tunnel surface cracks based on image recognition according to claim 4, wherein in the step 3, the construction process of the inverse motion compensation model involves converting the estimated rotation matrix and translation vector into a transformation operator of an image coordinate system; calculating displacement difference values between poses at adjacent moments aiming at transient translational blurring caused by vehicle running vibration, constructing a degradation function representing a motion blurring kernel function in a frequency domain when pixel displacement corresponding to the displacement difference values in exposure time exceeds a preset pixel threshold value, and deconvolving an image by using an inverse filtering algorithm; Aiming at perspective distortion generated by the fact that a camera optical axis is not perpendicular to a tunnel wall surface, projecting a current inclined imaging plane onto a reference plane parallel to a tunnel lining surface by utilizing a homography matrix, and remapping each pixel coordinate in an image; When the geometric correction is executed, the system dynamically corrects the depth factor in the projective transformation matrix according to the object distance data fed back by the laser ranging unit in real time, so as to ensure that the actual physical length represented by each pixel in the finally generated standardized image is kept constant when the detection platform is transversely offset.
  6. 6. The method for detecting tunnel surface cracks based on image recognition according to claim 5, wherein the generation process of the standardized tunnel surface image further comprises correction of intrinsic distortion of a lens, a coordinate offset correction mapping table covering a full pixel area is established by means of a radial distortion coefficient and a tangential distortion coefficient calibrated in advance, sub-pixel level resampling is performed on each pixel position, and stretching phenomenon generated in an image edge area is eliminated; Step 3 also adopts an image reprojection technology based on a non-parametric model, builds a dense lookup table containing nonlinear distortion in the camera in advance through a large field-of-view calibration field, and carries out local affine transformation adjustment on each image subarea according to the camera imaging model when processing pose increment; In the data processing level, the method relates to a multithreading parallel processing architecture, an image acquisition task, a pose estimation task, a crack identification task and a global positioning task are distributed to different cores of a computing unit, and exchange of data flows is realized through a high-speed memory sharing mechanism and a double-buffer mechanism.
  7. 7. The method for detecting the surface cracks of the tunnel based on the image recognition according to claim 6, wherein in the step 4, a preset crack detection neural network model adopts a symmetrical structure of an encoder and a decoder, wherein the encoder part sequentially compresses the space dimension of the feature map and increases the depth of the feature channel by cascade multi-layer convolution operation, normalization operation and nonlinear activation operation, and extracts deep semantic information of the crack in the image; the decoder part gradually restores the spatial resolution of the image through up-sampling operation, and utilizes a jump connection structure to fuse the low-layer high-resolution characteristic in the encoder with the high-layer semantic characteristic in the decoder; At the tail end of the network structure, a space and channel attention mechanism module is introduced, importance weights of all areas of the image are automatically learned, the importance weights are focused on a pixel set with crack characteristics, and false detection caused by water seepage, oil stains and construction joint background interference in a tunnel is restrained; the crack detection neural network model also introduces a knowledge distillation technology, and the crack discrimination knowledge learned by the expert model is transferred into the lightweight student model, so that the processing time of a single frame image is less than twenty milliseconds.
  8. 8. The method for detecting the surface cracks of the tunnel based on image recognition according to claim 7, wherein the geometric outline feature extraction of the cracks comprises morphological skeletonizing processing of a crack mask image, obtaining a central axis of the crack through a continuous stripping algorithm, searching edge pixel points to two sides along the normal direction of the central axis, calculating Euclidean distances between the two pairs of edge pixel points to obtain the width of the crack, and calculating the maximum width, the average width and the total length accumulated by the central axis; all the measurement values based on the pixel scale are converted into millimeter-level values with physical significance by combining internal parameter matrix data, distortion correction coefficients and real-time feedback object distance data of a camera; The method further comprises an online updating strategy of the deep learning model, when the system encounters a novel disease or a special background texture, the artificial intervention unit confirms a suspected region, a marked new sample is added to a training set, and model weights in the vehicle-mounted calculation unit are updated through an incremental learning algorithm; and in the post-processing stage of mask generation, adding a classifier based on physical priori, and dividing the identified target area into structural stress cracks, surface shrinkage cracks, water seepage trace interference and construction scratches by utilizing the geometric aspect ratio characteristics, the direction distribution characteristics and the relative position relation between the cracks and lining joints.
  9. 9. The method for detecting the surface cracks of the tunnel based on the image recognition according to claim 8, wherein in the step 5, the global world coordinate system is established based on a preset reference point at the entrance of the tunnel, and the synchronous positioning and mapping algorithm continuously maintains and dynamically updates a sparse point cloud map recording the coordinates of the space points of the remarkable geometric features and the descriptors of the feature vectors in the running process; When the detection vehicle passes through the same area again, the association between the current frame and the map points is realized through descriptor matching, a loop detection mechanism is triggered, the loop detection mechanism converts image features into visual word vectors by adopting a word bag model, and whether the detection platform passes through the visited position or not is identified by calculating the similarity between the current vector and the historical image vectors; once loop is detected, the system starts a global pose graph optimization algorithm, takes poses of all historical nodes as variables to be optimized, and takes closed-loop constraint as an error item to adjust a global track so as to minimize global errors; The sparse point cloud map is simplified by voxel grid filtering, service life attenuation evaluation is carried out on old characteristic points in the point cloud map, and the characteristic points which are stably observed in multiple inspection are reserved.
  10. 10. The method for detecting the surface cracks of the tunnel based on image recognition according to claim 9, wherein the history evolution monitoring process comprises automatic alignment logic, a system retrieves history images and history crack characteristics in the same physical coordinate range from a database based on a spatial index structure, and the database stores detection data in a segmented mode according to the number of the mileage of the tunnel and the number of the ring patch; Registering the current crack mask point cloud and the historical crack mask point cloud by utilizing an improved iterative nearest point algorithm, and automatically generating a tunnel structure health diagnosis report by calculating the area growth percentage, the width expansion quantity, the length expansion quantity and the newly-added branch condition of the same crack under different periods; If the crack variation exceeds a preset safety threshold, automatically triggering a grading alarm instruction by the system; The method is applied to an automatic tunnel inspection robot, the automatic tunnel inspection robot has the capability of autonomously planning a path, when a suspected severe crack area is detected, the travelling speed is automatically reduced or a stay shooting instruction is executed at a preset coordinate position, and a long exposure is utilized to acquire an image with high signal to noise ratio; And finally, constructing a three-dimensional visualized digital twin tunnel model by the system, and displaying the crack images of historical years, physical dimension parameters, environmental temperature and humidity records and predicted evolution curves of crack development.

Description

Tunnel surface crack real-time detection method based on image recognition Technical Field The invention belongs to the technical field of image data processing and structure detection, and particularly relates to a tunnel surface crack real-time detection method based on image recognition. Background With the continuous expansion of the scale of traffic infrastructure, long-term health monitoring of tunnel structures has become an important component for ensuring traffic safety. The cracks on the surface of the tunnel lining are used as key indexes for evaluating the stability and durability of the structure, and the detection accuracy and timeliness of the cracks are related to the scientificity of the later maintenance decision. In recent years, high-speed detection vehicles are used for carrying high-resolution imaging equipment for automatic inspection, so that the conventional manual visual inspection mode is gradually replaced, and the method becomes a main technical means for large-scale and high-efficiency tunnel defect inspection. The crack identification technology based on computer vision realizes automatic positioning and geometric parameter quantification of the micro cracks by collecting continuous tunnel wall images and utilizing a deep learning or feature extraction algorithm. Under the actual operation scene, in order to finish inspection under the premise of not interfering traffic, a detection system is usually required to be installed on a platform moving at a high speed, and full coverage scanning of the whole line surface of a tunnel is realized by integrating an image acquisition module and a positioning module. This process requires that the system be able to maintain sampling frequency and positioning accuracy during dynamic travel to ensure that every disease feature is clearly captured and that accurate spatial attributes are assigned. In the prior art, engineering problems are faced in the high-speed detection process, and the traditional image acquisition scheme is extremely easy to be influenced by track irregularity or mechanical vibration of a vehicle, so that motion blur and geometric distortion exist in an original image captured by a camera. The positioning logic of the existing system mainly depends on a single odometer or simple positioning equipment, and under a long-distance and closed tunnel environment, accumulated errors can cause the physical coordinate offset of cracks to reach the order of meters, so that high-precision global space correlation is difficult to establish. The existing detection means lack a dynamic sensing and compensation mechanism for the micro pose change of a camera, and sub-centimeter level pixel alignment can not be carried out on crack images with different detection periods, so that the reliability of the crack images in the process of monitoring the crack evolution trend and the structural degradation is reduced. Disclosure of Invention The invention aims to provide a tunnel surface crack real-time detection method based on image recognition, which can solve the problems in the background technology. In order to achieve the purpose, the technical scheme adopted by the invention is that the tunnel surface crack real-time detection method based on image recognition comprises the following specific steps: Step 1, constructing a multi-sensor fusion dynamic sensing system, and synchronously triggering a data acquisition instruction in the moving process of a detection platform by utilizing a high-resolution imaging unit, an inertia measurement unit and a laser ranging unit which are arranged on the detection platform to acquire continuous original sequence images, instantaneous angular velocity information, instantaneous acceleration information and preset distance data of the detection platform relative to the inner wall of the tunnel; Step 2, executing a front-end mileage calculation method based on visual synchronous positioning and map construction, and estimating six-degree-of-freedom pose parameters of a detection platform in a tunnel local space coordinate system in real time by extracting characteristic points in an original sequence image and performing inter-frame matching and combining motion increment output by an inertia measurement unit, wherein the pose parameters comprise a three-dimensional rotation matrix and a three-dimensional translation vector; Step 3, implementing dynamic drift compensation and image de-distortion processing, calculating the tiny vibration displacement of the detection platform at the exposure moment by using the pose parameters estimated in the step 2, constructing a reverse motion compensation model, performing pixel-level resampling and geometric correction on the original sequence image, eliminating motion blur and perspective distortion generated by shake of the detection platform, and generating a high-definition standardized tunnel surface image; Step 4, executing depth semantic segmentation and