A metro tunnel unmanned inspection method and system based on images

Abstract

The application discloses a kind of subway tunnel unmanned inspection method and system based on image, belong to city rail transit infrastructure maintenance technical field.The method includes the following steps: tunnel image and mileage data are synchronously acquired, and multi-dimensional observation evidence field vector drives state machine to jump in main state, auxiliary state and recovery state;Main state matches absolute assembly topological phase coordinates to execute microcosmic pose locking and update manifold control points;Auxiliary state fuses vision and mileage data to update coordinate track and cache temporary normalized image;Recovery state is corrected by back-priming resampling generation normalized image according to jump deviation through pose graph optimization to calculated pose;Finally, image difference processing is carried out, and system state derived artifact is removed, and effective surface disease evolution data is output.The application is processed through multi-state cooperation and data closed loop, effectively suppresses the cumulative mileage drift error of long distance inspection, guarantees that space coordinates and image sequence are strictly aligned, and the accuracy of disease data extraction is improved.

Description

Technical Field

[0001] This invention relates to the field of urban rail transit infrastructure maintenance technology, specifically to an image-based unmanned inspection method and system for subway tunnels. Background Technology

[0002] Subway tunnels are typical environments without external positioning signals, and existing unmanned inspection equipment typically relies on wheeled odometers or inertial sensors for dead reckoning. During long-distance continuous inspections, a single kinematic calculation can generate cumulative mileage drift errors that are difficult to eliminate. To correct these errors, existing technologies often use image matching to locate tunnel segments by extracting structural features such as joints or bolt holes. However, due to objective assembly tolerances in the production and assembly of tunnel segments, and the macroscopic deformation of the overall physical cross-section of the tunnel after long-term service, traditional rigid feature matching methods are difficult to adapt to these minute differences and physical deformations, leading to absolute coordinate locking failures or limited matching accuracy.

[0003] On the other hand, when visual features are obscured by dirt, causing temporary anomalies in trajectory estimation, existing inspection methods lack a post-processing correction mechanism for historical trajectories during periods of lockout. This results in a lack of strict spatial correspondence between the generated current inspection image sequence and the baseline image. Under these circumstances, if the two images are directly compared differentially, misaligned normal joints and hole edges of pipe segments are easily misjudged by the algorithm as newly added structural defects, thus generating a large number of differential artifacts in the data layer. This increases the cost of manual verification and seriously affects the accuracy of extracting true surface defect evolution data.

[0004] Therefore, this invention proposes an image-based unmanned inspection method and system for subway tunnels to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an image-based unmanned inspection method and system for subway tunnels, which solves the problems of spatial misalignment of isomorphic images and artifacts generated by differential extraction of surface defects caused by cumulative mileage drift, differences in tunnel segment structure, and lack of backward correction for local lockouts in unmanned subway tunnel inspections.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an image-based unmanned inspection method for subway tunnels, comprising the following steps: Simultaneously acquire two-dimensional image frames and mileage data of the tunnel, extract multiple image feature parameters in the dynamic mileage search domain to construct a multidimensional observation evidence field vector, and drive the state machine to transition between the main state, auxiliary state and recovery state; In the main state, the absolute assembly topology phase coordinates are matched, the mixed high-frequency sequence is extracted to perform elastic phase-locked path matching and micro pose locking, and the manifold control point update and anomaly isolation update are performed. When in the auxiliary state, the optimal displacement increment is calculated to update the kinematic coordinate orbit, the theoretical topology prediction phase is extrapolated to maintain the topology prediction coordinate orbit, and a temporary normalized image is generated by combining the kinematic coordinate orbit, the topology prediction coordinate orbit and the historical manifold control point set. When in the recovery state, perform topology prediction coordinate trajectory consistency check to generate jump deviation, perform pose error optimization and refeedback and mesh expansion resampling based on jump deviation, generate normalized image or truncate optimization loop to trigger reference repositioning; A source confidence label matrix for standardized images is established and differential candidate regions are extracted. Attribution determination is performed by combining the differential candidate regions with the source confidence label matrix, system state-derived artifacts are filtered, and effective surface disease evolution data are output.

[0007] Preferably, the process involves simultaneously acquiring two-dimensional tunnel image frames and mileage data, extracting multiple image feature parameters within the dynamic mileage search domain to construct a multidimensional observation evidence field vector, and driving the state machine to transition between the dominant state, auxiliary state, and recovery state, including: A dynamic mileage search domain for generating tunnel two-dimensional image frames is generated based on preset forward slip tolerance, preset short-term reverse disturbance tolerance, and mileage data. Within the dynamic mileage search domain, multiple image feature parameters of the tunnel two-dimensional image frame are extracted, including gray-scale projection valley significance coefficient, local texture information entropy, anisotropic response parameter and unidirectional phase consistency parameter. The multiple image feature parameters are normalized and spliced ​​to generate a multidimensional observation evidence field vector. The multidimensional observation evidence field vector is compared with the preset observation condition matrix, and a state transition determination is performed in combination with the current system state. The observation condition matrix is ​​composed of the confidence requirement vectors of the main state, auxiliary state, and recovery state.

[0008] Preferably, matching the absolute assembly topology phase coordinates includes: Using edge detection operators, Hough linear transform and Hough circular transform, the intersection points of annular joints, longitudinal joints, bolt holes and cross joints are extracted from the tunnel 2D image frame as structural primitives and defined as nodes. The relative spatial distance and angular relationship between the extracted nodes are defined as edges, and a directed observation topology segment is established. The differences in node type, side length Euclidean distance, and included angle cosine value between the directed observation topology fragment and the global design diagram structure in the prior benchmark database are calculated. Independent weights are assigned to the differences in node type, side length Euclidean distance, and included angle cosine value, and a weighted sum is performed to generate a comprehensive topology similarity score. The position coordinates with the highest comprehensive topology similarity score that are greater than the set judgment threshold are set as the absolute assembly topology phase coordinates.

[0009] Preferably, extracting the mixed high-frequency sequence to perform elastic phase-locked path matching and micro-pose locking includes: Multi-resolution wavelet decomposition is performed on the geometric spatial signal of the physical edge of the tunnel two-dimensional image frame. Low-frequency approximation coefficients are discarded and high-frequency detail coefficients are reconstructed. High-frequency deviation signals reflecting the small differences of the tunnel segments are extracted to generate a hybrid high-frequency sequence. Based on the multidimensional observation evidence field vector, absolute assembly topology phase coordinates, and historical topology prediction deviation, a soft constraint weight window is generated. Under the guidance of the soft constraint weight window, a dynamic warping algorithm is used to calculate the cumulative distance matrix between the mixed high-frequency sequence and the benchmark period high-frequency assembly tolerance sequence. The elastic phase-locked path is extracted by introducing step size constraint, and the elastic phase-locked residual is calculated based on the elastic phase-locked path. Within the local neighborhood determined by the elastic phase-locked path, a one-dimensional normalized cross-correlation matching is performed on the hybrid high-frequency sequence and the reference period high-frequency assembly tolerance sequence to complete the locking of the micro pose.

[0010] Preferably, performing manifold control point updates and anomaly isolation updates includes: Spatial back projection is performed on the intersection of circumferential joints, longitudinal joints, bolt holes and cross joints to obtain the current three-dimensional anchor point set. The Euclidean spatial distance between each anchor point in the current three-dimensional anchor point set and the prior design section depth data is calculated, and the Euclidean spatial distance is used as the spatial deviation of each anchor point. Extract all anchor points with spatial deviations less than or equal to the preset deformation tolerance threshold, generate local anchor point residual regions, use B-spline surfaces to absorb the macroscopic low-frequency deformation of the overall tunnel cross section, introduce thin plate spline energy minimization constraints in local anchor point residual regions, form a local residual compensation field superimposed on the B-spline surface, and update the corresponding manifold control points in the manifold memory bank according to the local residual compensation field. Extract anchor points whose spatial deviation is greater than the preset deformation tolerance threshold, and write the anchor points into the macroscopic deformation anomaly flag in the manifold memory.

[0011] Preferably, the optimal displacement increment is calculated to update the kinematic coordinate trajectory, the theoretical topological prediction phase is extrapolated to maintain the topological prediction coordinate trajectory, and a temporary normalized image is generated by combining the kinematic coordinate trajectory, the topological prediction coordinate trajectory, and the historical manifold control point set, including: Extract multiple sets of tunnel 2D image frames acquired during the system's auxiliary state, record two adjacent sets of tunnel 2D image frames as adjacent frames, calculate the polynomial expansion coefficients of adjacent frames, compare the changes in the polynomial coefficients of adjacent frames to obtain 2D pixel motion vectors, combine the 2D pixel motion vectors with the camera intrinsic parameter matrix to generate 3D spatial displacement increments, and record the 3D spatial displacement increments as visual mileage displacement observation vectors. The observation residual between the visual odometer displacement observation vector and the wheel odometer physical displacement observation vector is calculated using the chi-square test. The visual odometer displacement observation vector and the wheel odometer physical displacement observation vector are weighted and fused using a Kalman filter framework to output the optimal displacement increment. The kinematic coordinate trajectory is updated based on the optimal displacement increment of all adjacent frames and the current camera pose. Extract the initial topological phase before entering the auxiliary state, combine the standard assembly topological cycle length of the tunnel segment, the current linear velocity and current acceleration of the vehicle, calculate the theoretical topological prediction phase of each tunnel two-dimensional image frame in the continuous auxiliary state frame based on the uniform acceleration kinematic model, and generate the topological prediction coordinate trajectory by combining the theoretical topological prediction phases of all tunnel two-dimensional image frames. Based on the theoretical topological prediction phase of the tunnel 2D image frame, the historical manifold control point set corresponding to the location is retrieved from the manifold memory. The optimal displacement increment of the tunnel 2D image frame is mapped to the theoretical topological prediction phase interval determined by the topological prediction coordinate track to generate the corresponding comprehensive estimated camera pose. The temporary normalized image is generated by combining the comprehensive estimated camera pose with the historical manifold control point set.

[0012] Preferably, the process involves performing a topology prediction coordinate trajectory consistency check to generate a jump deviation, and then performing pose error optimization refeedback and mesh unfolding resampling based on the jump deviation to generate a normalized image or truncate an optimized loop to trigger reference repositioning, including: Extract the most recent auxiliary state's comprehensive inference camera pose, calculate the jump deviation between the comprehensive inference camera pose and the absolute assembly topology phase coordinates, and update the historical topology prediction deviation based on this jump deviation. If the jump deviation is less than or equal to the jump tolerance threshold, extract all the comprehensive inferred camera poses during the auxiliary state, define each comprehensive inferred camera pose as a state node, define the relative pose transformation vector between adjacent state nodes as a constraint edge, and construct a pose graph optimization framework by combining the state nodes and constraint edges; perform pose graph optimization based on the pose graph optimization framework to correct the comprehensive inferred camera pose; extract a temporary normalized image with auxiliary state extrapolation labels, perform a three-dimensional ray intersection operation on the temporary normalized image according to the corrected comprehensive inferred camera pose to obtain the three-dimensional spatial intersection points, and resample the three-dimensional spatial intersection points according to the two-dimensional orthogonal mesh interpolation with a fixed physical resolution to generate a normalized image; If the jump deviation exceeds the jump tolerance threshold, the reverse allocation loop of the pose graph optimization is truncated, the comprehensive inference of the camera pose during the auxiliary state is not corrected, blind zone trajectory loss labels are added to the temporary normalized images in the data buffer, and the absolute pose of the camera is re-bound using the absolute assembly topology phase coordinates to achieve reference repositioning.

[0013] Preferably, the process includes establishing a source confidence label matrix for the normalized image and extracting differential candidate regions, including: Based on the orthogonal isomorphic grid coordinate system, a source confidence label matrix containing Boolean flags of the system state is constructed pixel by pixel; Using the baseline isomorphic tunnel two-dimensional image frame as a reference template, the normalized image is subjected to histogram specification processing; In the orthogonal isomorphic grid coordinate system, the gray-level absolute difference matrix between the reference period isomorphic tunnel two-dimensional image frame and the normalized image after histogram specification is calculated, and the gray-level absolute difference matrix is ​​binarized by setting the gray-level difference threshold. The two-scan algorithm based on eight-connected neighborhoods extracts independent connected regions from the binary gray-level absolute difference matrix and outputs differential candidate regions.

[0014] Preferably, attribution determination is performed by combining the differential candidate region and the source confidence label matrix, filtering system state-derived artifacts, and outputting valid surface defect evolution data, including: Extract the grid coordinates of all foreground pixels within the differential candidate region to construct a two-dimensional covariance matrix. Obtain the eigenvector corresponding to the largest eigenvalue through eigenvalue decomposition as the main extension direction vector. Extract the deformation direction vector of the system operation and calculate the directional consistency coefficient between the main extension direction vector and the deformation direction vector. If the directional consistency coefficient of the differential candidate region is greater than the directional determination threshold, and the corresponding source confidence label matrix contains a resampling boundary flag, a macroscopic deformation anomaly flag, an elastic phase-locked loop residual anomaly flag, or a blind zone trajectory loss flag, then the differential candidate region is determined to be a system state-derived artifact and is removed; otherwise, the differential candidate region is determined to be valid surface disease evolution data and is output.

[0015] This invention also provides an image-based unmanned inspection system for subway tunnels, comprising: The state control module is configured to synchronously acquire tunnel two-dimensional image frames and mileage data, generate a dynamic mileage search domain and output a multi-dimensional observation evidence field vector, and drive the state machine to transition between the main state, auxiliary state and recovery state. The dominant manifold module is configured to match the absolute assembly topology phase coordinates, extract the mixed high-frequency sequence to perform microscopic pose locking, and update the manifold control points when it is in the dominant state. The auxiliary state cache module is configured to, when in auxiliary state, calculate the optimal displacement increment to update the kinematic coordinate trajectory, extrapolate the theoretical topology prediction phase to maintain the topology prediction coordinate trajectory, and generate a temporary normalized image and store it in the data cache area. When the recovery module is configured to be in recovery mode, it calculates the jump deviation and generates a normalized image or triggers a reference repositioning based on the jump deviation. The differential attribution module is configured to generate a source confidence label matrix and differential candidate regions for normalized images. It combines the source confidence label matrix and differential candidate regions to generate directional consistency coefficients, eliminates system state-derived artifacts based on the directional consistency coefficients, and outputs effective surface disease evolution data.

[0016] The image-based unmanned inspection method and system for subway tunnels provided by this invention have the following beneficial effects: 1. The method provided by this invention drives a state machine to adaptively transition between the main state, auxiliary state, and recovery state by extracting multi-dimensional observation evidence field vectors. In the main state, microscopic pose locking is performed; in the auxiliary state, the calculated trajectory is maintained and temporary images are cached; and then in the recovery state, pose map optimization and backfeed correction are performed using the jump deviation of the absolute topological phase. This multi-state collaborative operation and data closed-loop processing mechanism can effectively suppress the cumulative mileage drift error caused by long-distance inspections in tunnel environments lacking continuous external positioning signals, ensuring strict alignment of spatial coordinates and image sequences. Combined with a differential filtering mechanism, it improves the objectivity and accuracy of surface disease evolution data extraction. 2. This invention extracts high-frequency deviation signals by performing multi-resolution wavelet decomposition on the geometric spatial signals of the physical edges of images, and then uses a dynamic warping algorithm guided by a soft-constraint weighted window to extract the elastic phase-locked path. This mechanism can adapt to the minute structural differences of the tunnel segments, avoid the problem of easy failure of conventional rigid template matching, and enable the system to reliably lock the microscopic absolute coordinates in the local neighborhood, thus improving the robustness of the benchmark positioning. 3. This invention utilizes the spatial deviation between the three-dimensional anchor point set obtained by spatial back projection and the prior design section to construct a local residual compensation field superimposed on a B-spline surface that absorbs macroscopic low-frequency deformation. This compensation mechanism can dynamically adjust the manifold control points according to the actual spatial deformation, so that the resampled normalized image can truly reflect the current three-dimensional topological surface of the tunnel, reducing image difference misjudgment caused by physical deformation; 4. This invention constructs a pose graph optimization framework in the recovery state that integrates multi-dimensional observation weights such as optical flow confidence and odometry consistency residuals. When the jump deviation is within the allowable range, the Levenberg-Marquardt algorithm is used to jointly iteratively optimize the estimated camera pose. This backfeeding correction mechanism realizes backward smoothing of the historical blind zone trajectory and ensures the physical spatial consistency of the temporarily stored images during the auxiliary state during the final resampling. Attached Figure Description

[0017] Figure 1 This is an architecture diagram of an image-based unmanned subway tunnel inspection system according to an embodiment of the present invention; Figure 2 This is a flowchart of an image-based unmanned inspection method for subway tunnels according to an embodiment of the present invention; Figure 3 A flowchart for generating a multidimensional observation evidence field vector and state switching in step S100 of this embodiment of the invention; Figure 4 This is a flowchart of step S200, the determination of main state locking and physical deformation decoupling, in an embodiment of the present invention. Figure 5 This is a flowchart of step S300, auxiliary state dual-track calculation and blind zone temporary storage, in an embodiment of the present invention. Figure 6 This is a flowchart of step S400, the recovery state reverse pose graph optimization and refeedback mechanism, in an embodiment of the present invention. Figure 7 This is a flowchart of step S500, source label construction and artifact differential attribution determination, in an embodiment of the present invention. Figure 8 This is a schematic diagram of the curve comparison of the blind zone three-dimensional pose estimation trajectory in an embodiment of the present invention. Figure 9 This is a scatter plot of the attribute distribution of the differential candidate region in an embodiment of the present invention.

[0018] The reference numerals in the attached figures are: State Control Module 10, Main State Manifold Module 20, Secondary State Cache Module 30, Recovery and Backfeed Module 40, and Differential Attribution Module 50. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0020] Reference Figure 1 The present invention provides an image-based unmanned inspection system for subway tunnels, comprising a state control module 10, a main state manifold module 20, an auxiliary state cache module 30, a recovery and recharge module 40, and a differential attribution module 50.

[0021] The state control module 10 is configured to synchronously acquire tunnel two-dimensional image frames and mileage data, extract multiple image feature parameters and generate a multidimensional observation evidence field vector, and determine whether the system enters the main state, auxiliary state or recovery state based on the multidimensional observation evidence field vector.

[0022] The main state manifold module 20 is configured to match the absolute assembly topology phase coordinates when the system is in the main state, extract the high-frequency tolerance sequence to perform micro pose locking, and update the manifold control points according to the spatial deviation and write them into the manifold memory.

[0023] The auxiliary state cache module 30 is configured to calculate the optimal displacement increment and update the kinematic coordinate trajectory when the system is in the auxiliary state, extrapolate the theoretical topology prediction phase to maintain the topology prediction coordinate trajectory, and generate a temporary normalized image by combining the kinematic coordinate trajectory, the topology prediction coordinate trajectory and the historical manifold control point set retrieved from the manifold memory and cache it.

[0024] The recovery and recharge module 40 is configured to acquire the absolute assembly topology phase and perform consistency verification with the topology prediction coordinate orbit when the system is in the recovery state to generate a jump deviation, allocate the absolute drift error based on the jump deviation and resample to generate a normalized image, or truncate the error allocation loop to trigger the reference repositioning.

[0025] The differential attribution module 50 is configured to receive normalized images, establish a source confidence label matrix, extract differential candidate regions from the normalized images, input the source confidence label matrix and differential candidate regions into a rule-based classifier based on geometric orientation consistency to perform attribution determination, filter artifacts, and output valid surface disease evolution data. Example 2

[0026] Reference Figure 2 This invention provides an image-based unmanned inspection method for subway tunnels, comprising the following steps: S100 synchronously acquires tunnel two-dimensional image frames and mileage data, constructs a dynamic mileage search domain, extracts image feature parameters based on the dynamic mileage search domain to generate a multi-dimensional observation evidence field vector, and controls the system to transition to the main state, auxiliary state or recovery state based on the multi-dimensional observation evidence field vector. S200 matches the absolute assembly topology phase coordinates in the main state, uses high-frequency tolerance sequences to achieve microscopic pose locking, and updates manifold control points based on spatial deviations. S300, in the auxiliary state, calculate the optimal displacement increment to update the kinematic coordinate trajectory, extrapolate the theoretical topology prediction phase to maintain the topology prediction coordinate trajectory, and combine the kinematic coordinate trajectory, the topology prediction coordinate trajectory and the historical manifold control point set to generate a temporary normalized image. S400 performs topology prediction coordinate trajectory consistency verification in the recovered state to generate jump deviation. Based on the jump deviation, it performs pose error optimization and refeedback and mesh expansion resampling to generate normalized images, or truncates the optimization loop to trigger reference repositioning. S500 establishes a source confidence label matrix for standardized images and extracts differential candidate regions. It then combines the differential candidate regions with the source confidence label matrix to perform attribution determination, filters system state-derived artifacts, and outputs valid surface disease evolution data.

[0027] To further clarify the implementation of each technical aspect of the present invention, the following will provide a detailed description of each functional module involved in the implementation of Embodiment 1 and its internal processing flow.

[0028] Reference Figure 3 In the method provided by this invention, step S100 specifically involves acquiring two-dimensional tunnel image frames and mileage data, and constructing a multi-dimensional observation evidence field vector. The state control module 10 is configured to perform temporal correlation between the two-dimensional tunnel image frames and mileage data, and to construct a multi-dimensional observation evidence field to drive stable transitions in the state machine of the image-based unmanned subway tunnel inspection system. Specifically, it includes the following steps: S101, based on the sampling time of the wheel odometer, the camera exposure timestamp, and a unified time reference, completes the temporal correlation between the tunnel two-dimensional image frames and mileage data, generating a dynamic mileage search domain. Specifically, in this embodiment, the state control module 10 calls the prior design section depth data corresponding to the tunnel section to be inspected in the prior reference database as the initial geometric reference, according to the preset positive slip tolerance k. f and short-term reverse disturbance tolerance k b Set the dynamic mileage search domain for the current frame [S] lower ,S upper [This is to accommodate instantaneous reverse fluctuations in mileage caused by wheel-rail contact, transmission clearance, or local vibration.] upper S lower These are the upper and lower limits of the dynamic mileage search domain, S. upper =S c +v·Δt·(1+k f ), S lower =S c +v·Δt·(1-k b In the formula, S c The reference mileage of the previous valid frame is represented by v; the current linear velocity output by the wheel odometer is represented by v; and Δt represents the time interval between the current frame and the previous valid frame. The current frame refers to the tunnel 2D image frame acquired at the current moment. The previous valid frame refers to the most recently acquired valid tunnel 2D image frame before the current frame. As a preferred embodiment, the positive slip tolerance k... f The value is obtained through statistical calibration of track friction coefficient and train acceleration / deceleration performance test data, and its range is typically set to 0.05 to 0.15. The short-time reverse disturbance tolerance k... b The value is determined by prior statistical parameters of the transmission gear clearance, typically ranging from 0.01 to 0.05. These parameter settings ensure that the dynamic mileage search domain can cover the actual physical displacement range under preset slip and disturbance conditions.

[0029] The prior benchmark database is initialized and established before the system is first deployed to inspect the corresponding tunnel section. It establishes a joint index based on the line section, line mileage, segment ring number, and circumferential position. Under normal lighting conditions, unobstructed structural elements, and valid mileage data, it synchronously collects benchmark-period tunnel 2D image frames and mileage data based on a unified time reference to determine the absolute assembly topology phase coordinates of each benchmark-period tunnel 2D image frame. Then, based on the joint index, it associates and saves each benchmark-period tunnel 2D image frame with its corresponding absolute assembly topology phase coordinates, prior design section depth data, and design data of the tunnel section to be inspected. Further, it generates a global design diagram structure based on the design data of the tunnel section to be inspected, generates a benchmark-period high-frequency assembly tolerance sequence based on the benchmark-period tunnel 2D image frames and absolute assembly topology phase coordinates, and generates benchmark-period isomorphic tunnel 2D image frames by combining the benchmark-period tunnel 2D image frames, absolute assembly topology phase coordinates, prior design section depth data, and camera calibration parameters. Finally, it writes the global design diagram structure, benchmark-period high-frequency assembly tolerance sequence, and benchmark-period isomorphic tunnel 2D image frames into the prior benchmark database according to the joint index. It should be noted that for baseline tunnel 2D image frames with incomplete structural primitives, abnormal exposure, or invalid mileage data, binding and data entry will not be performed.

[0030] S102, within the dynamic mileage search domain, extract multiple image feature parameters of the current frame to construct a multidimensional observation evidence field vector. The image feature parameters include grayscale projection valley significance coefficient, local texture information entropy, anisotropic response parameters, and unidirectional phase consistency parameters.

[0031] The significance coefficient of the grayscale projection valley is calculated. Specifically, the state control module 10 performs grayscale integral projection of the current frame along the circumferential direction of the tunnel to generate a one-dimensional projection curve. The local minima and local maxima of this curve are extracted within a defined neighborhood. The local minima characterize the dark stripe features of the tunnel joints. Further, the local minima are denoted as the valley depth P. min The local maxima are denoted as the background reference P. max Based on the valley depth P min and background benchmark P max Calculate the significance coefficient C of grayscale projection valleys. valley C valley =(P max -P min ) / (P max +ε0). Where ε0 is a minimal positive real number bias term to prevent division by zero, for example, a value of 10. -6 .

[0032] Calculate the local texture information entropy. Specifically, the state control module 10 extracts the grayscale values ​​of all pixels in the current frame and generates a grayscale value matrix. Based on the set pixel statistical distance, it extracts the grayscale co-occurrence matrix H of the grayscale value matrix at the four basic direction angles. gd After normalization, the corresponding probability matrix P is obtained. gd , gd=1,2,3,4. As a preferred method, the pixel statistical distance is set to 1 to 5 pixels, and the four basic orientation angles are 0°, 45°, 90°, and 135°. Subsequently, based on the probability matrix P gd Calculate texture information entropy E gd , Among them, G v For example, when G... v When the value is 256, the grayscale values ​​i and j are integers in the range [0, 255]. gd (i,j) represents the probability matrix P gd The element in the i-th row and j-th column corresponds to the normalized probability of the gray-level pair (i,j). ε1 is a set minimum positive real number compensation term, for example, a value of 10. -6 Used to avoid when P gd An anomaly occurs in the logarithmic function calculation when (i,j)=0. Further, the texture information entropy E for the four basic orientation angles is calculated. gd The mean value is used to obtain the local texture information entropy E, where E = (E1 + E2 + E3 + E4) / 4. In this invention, the local texture information entropy E quantifies the degree of deviation of the texture distribution of the tunnel region corresponding to the current frame at the current moment from the normal texture distribution in the reference period. If the local texture information entropy is not within the preset normal texture range, it indicates that the tunnel region may currently be in a state of surface coverage, crystallization, water stains, reflection, or abnormal exposure.

[0033] The anisotropic response parameters are calculated. Specifically, since tunnel joints exhibit typical linear structural characteristics, the state control module 10 constructs a Hessian matrix using the second-order partial derivatives of the current frame pixels and performs eigenvalue decomposition to extract two eigenvalues ​​reflecting local curvature changes. These eigenvalues ​​are then sorted according to their absolute values, with the larger eigenvalue denoted as λ1 and the smaller eigenvalue as λ2. The anisotropic response parameter R is then calculated using a ratio model incorporating a smoothing factor: R = |λ1| / (|λ2|+ε2). Here, ε2 is a smoothing factor set to prevent the denominator from approaching 0, and its value is typically set to 10. -5 The larger the value of the anisotropic response parameter R, the stronger the gradient change of the current frame pixels in a certain direction and the gentler change in the orthogonal direction. In this case, the current frame has a clear linear structure, which means that the probability of tunnel seams in the corresponding region of the current frame is higher.

[0034] The unidirectional phase consistency parameter is calculated. Specifically, the state control module 10 performs frequency domain convolution response calculation on the current frame using a multi-scale Log-Gabor filter to obtain the complex response R at different frequency scales. n Based on complex response R n Generate the local amplitude A corresponding to frequency scale n n With local energy E n n=1,2,…,N f N f The total number of frequency scales is defined. The local amplitude A is calculated by combining all frequency scales. n With local energy E n Calculate the one-way phase coherence parameter PC. This allows us to obtain structural feature measurement data that is insensitive to changes in ambient lighting and local contrast attenuation. Among these, T... n ε is the energy compensation threshold set based on local noise. p To prevent the smoothing coefficient from being zero in the denominator, a value of 10 is used. -5 .

[0035] The state control module 10 normalizes the aforementioned image feature parameters based on extreme value mapping to eliminate dimensional differences in heterogeneous data. Then, the normalized grayscale projection valley significance coefficient C is... valley The local texture information entropy E', anisotropic response parameter R', and unidirectional phase consistency parameter PC' are multidimensionally spliced ​​together to generate a multidimensional observation evidence field vector O. t O t =[C valley ',E',R',PC'].

[0036] S103: The multidimensional observation evidence field vector of the current frame is compared with the preset observation condition matrix, and the state machine transition control logic is executed. The preset observation condition matrix includes confidence requirement vectors for the dominant state, auxiliary state, and recovery state. Each confidence requirement vector consists of feature confidence requirements for multiple image feature parameters, including valley significance requirements, texture entropy normal range requirements, anisotropic response requirements, and unidirectional phase consistency requirements. For the same image feature parameter, the feature confidence requirements differ in different states. Historical image datasets acquired under normal lighting conditions during the baseline period are extracted and a statistical model is established. Based on this statistical model, feature confidence requirements for each state and each image feature parameter are generated. As a preferred approach, to improve the consistency of system state assessment, for each state, the system extracts multiple sets of image feature parameter data for that state from the historical image dataset, calculates the mean and standard deviation of the multiple sets of data for each image feature parameter, and sets the corresponding feature confidence requirements according to the 3σ criterion, thereby obtaining the confidence requirement vector for that state.

[0037] The state control module 10 will control the multidimensional observation evidence field vector O of the current frame. t The observation quality of the current frame is evaluated by comparing the vectors with the confidence requirements of the main state, auxiliary state, and recovery state. In actual control, to avoid frequent state jitter caused by sudden changes in a single image feature parameter, the state control module 10 uses the multidimensional observation evidence field vector O of the current frame. t A comprehensive judgment is performed by combining the current system state and the feature consistency parameters of consecutive frames.

[0038] The continuous frames refer to the current frame extracted within a preset time window and several preceding tunneling two-dimensional image frames (e.g., 3 to 5 frames). A feature consistency parameter F is generated based on multiple multidimensional observation evidence field vectors corresponding to the continuous frames. con Specifically, multiple continuous values ​​of the same image feature parameter are extracted from the multiple multidimensional observation evidence field vectors, and the fluctuation variance of each image feature parameter within the time window is calculated. If the fluctuation variance of all image feature parameters is less than or equal to a preset stability threshold, then let F... con =1 indicates that the system can stably and continuously extract observation features that meet the quality standards; otherwise, let F = 1. con =0.

[0039] The execution logic of the comprehensive judgment is as follows: when the state control module 10 determines the multidimensional observation evidence field vector O of the current frame... t The confidence requirement vector O that satisfies the dominant state a When the system is currently in a recovery state or a dominant state, the control state machine transitions to the dominant state or maintains the dominant state; when the state control module 10 determines the multidimensional observation evidence field vector O of the current frame... t The confidence requirement vector O for the auxiliary state is satisfied c Furthermore, when local texture information entropy deviates from the normal texture range and anisotropic response parameters decrease due to illumination degradation or local occlusion, the control state machine switches to an auxiliary state; when the state control module 10 determines the multidimensional observation evidence field vector O of the current frame... t The confidence requirement vector O for the recovered state is satisfied. b Furthermore, the current system state is in an auxiliary state, and the feature consistency parameter F of consecutive frames... con When the value is 1, the control state machine switches to the recovery state. Preferably, to prevent frequent state reversals in the state boundary region, a minimum frame period T is preset. min The duration of the current state of the statistical system is measured in frames T. sys If T sys <T min The system maintains its current state. The minimum frame period T min It can be set to 10 to 15 frames.

[0040] Reference Figure 4 In the method provided by this invention, step S200 is specifically implemented as master-state absolute locking and manifold control point update. When the system transitions to the master state, the master-state manifold module 20 is configured to perform spatial absolute registration and physical reference update, specifically including the following steps: S201, extract structural primitives in the current frame and establish observation topology fragments to confirm the absolute assembly topology phase coordinates of the current frame. In this embodiment, the dominant manifold module 20 extracts the intersection points of circumferential seams, longitudinal seams, bolt holes, and cross seams in the current frame as key structural primitives. To accurately extract the edges and contours of the above structural primitives, the dominant manifold module 20 applies the Canny edge detection operator to Gaussian smooth the current frame to filter out high-frequency noise and calculates the gradient magnitude and direction of the pixels in the current frame. After non-maximum suppression and double-threshold edge connection, a continuous single-pixel wide edge image is obtained. Based on this, the dominant manifold module 20 uses Hough linear transform and Hough circular transform to map the edge pixels in the single-pixel wide edge image space to the parameter space. By detecting local peaks in the parameter space, it extracts the straight line segments representing seams (circumferential seams, longitudinal seams, and cross seams) and the circular primitives representing bolt holes.

[0041] The dominant manifold module 20 defines the extracted structural primitives as nodes, the lines connecting the nodes as edges, and the relative spatial distance and angular relationship between the nodes as attributes of the edges, thereby establishing a directed observation topology fragment representing local geometric features. Then, a subgraph isomorphism algorithm is used to calculate the differences between this directed observation topology fragment and the global design graph structure in the prior benchmark database in terms of corresponding node types, Euclidean distance of edge lengths, and cosine value of the included angle. Independent weights are assigned to these three difference dimensions, and a weighted sum is performed to generate a comprehensive topology similarity score. The global design graph structure is constructed based on the design data of the tunnel section to be inspected in the prior benchmark database. Specifically, the design positions of the intersections of circumferential joints, longitudinal joints, bolt holes, and cross joints are extracted from the design data. Each design position is used as a design node, and the distance and angular relationship between the design nodes are used as design edges. The global design graph structure is generated by combining the design nodes and design edges. As a preferred approach, the node type difference weight is set to 0.2, the edge length Euclidean distance difference weight is set to 0.4, and the included angle cosine value difference weight is set to 0.4.

[0042] Finally, the dominant manifold module 20 extracts the position coordinates with the highest comprehensive topological similarity score that is greater than a set judgment threshold, and outputs them as the absolute assembly topological phase coordinates of the current frame. The judgment threshold is set based on the comprehensive topological similarity scores of each tunnel region extracted from the baseline static scan data. Specifically, firstly, based on the joint index, the baseline tunnel two-dimensional image frames corresponding to all tunnel regions are extracted from the prior database. Based on the baseline tunnel two-dimensional image frames, directed observation topological fragments of each tunnel region in the baseline period are generated. The directed observation topological fragments of all tunnel regions in the baseline period are combined to generate baseline static scan data. Subsequently, the comprehensive topological similarity score between each directed observation topological fragment in the baseline static scan data and the global design diagram structure is calculated, and then the mean μ of all comprehensive topological similarity scores is calculated. top and standard deviation σ top Thus, the decision threshold T is obtained. top T top =μ top -q top ·σ top Among them, q top This is the confidence boundary coefficient, with a value ranging from 2 to 3. For example, in this embodiment, q top =3.

[0043] S202, based on the high-frequency tolerance sequence, performs topological constraint elastic micro-locking to eliminate pose bias caused by assembly tolerances and minor deformations. First, the dominant manifold module 20 uses one-dimensional discrete wavelet transform to perform multi-resolution decomposition of the geometric spatial signal of the physical edge. Specifically, the system uses Daubechies wavelets (e.g., db4 wavelet basis) to perform multi-level (usually 3 to 5 levels) decomposition of the spatial signal to obtain low-frequency approximation coefficients and high-frequency detail coefficients. The system discards the low-frequency approximation coefficients to filter out the low-frequency trend components representing the macroscopic geometric surface, and only reconstructs the high-frequency detail coefficients, thereby extracting the high-frequency deviation signal reflecting the minor differences in segment manufacturing and assembly, forming the current frame's hybrid high-frequency sequence.

[0044] Secondly, the dominant manifold module 20 generates a soft-constraint weight window based on the multidimensional observation evidence field vector, absolute assembly topology phase coordinates, and historical topology prediction bias output in step S100. As a preferred embodiment, this soft-constraint weight window exhibits a Gaussian distribution centered on the absolute assembly topology phase coordinates. Its attenuation variance is dynamically determined by the absolute value of the historical topology prediction bias, and its amplitude coefficient is determined by the multidimensional observation evidence field vector. The historical topology prediction bias is the jump bias from the most recent recovery state calculation. For example, the variance σ of the soft-constraint weight window... w Satisfy: σ w=σ0+η·|ΔT|, where σ0 is the variance of the base window, set according to the average interval of adjacent assembly tolerance peaks in the reference period tunnel two-dimensional image frame; η is the deviation amplification coefficient, preferably ranging from 0.5 to 2.0; both σ0 and η are positive real numbers; ΔT is the historical topology prediction deviation, and the larger |ΔT| is, the wider the set soft constraint weight window. Simultaneously, the confidence level of the observation evidence field is generated by combining the multidimensional observation evidence field vector and the historical topology prediction deviation. This confidence level is then scaled proportionally to obtain the amplitude coefficient of the soft constraint weight window. The confidence level of the observation evidence field is obtained by weighting the normalized grayscale projection valley significance, local texture information entropy, anisotropic response parameters, unidirectional phase consistency parameters, and historical topology prediction deviation confidence level; the values ​​of each weight coefficient γ1, γ2, γ3, γ4, and γ5 are all in the range [0,1], and satisfy γ1+γ2+γ3+γ4+γ5=1.

[0045] Subsequently, guided by the soft-constraint weight window, the dominant manifold module 20 extracts the corresponding high-frequency assembly tolerance sequence from the prior benchmark database based on the joint index of the current frame. It then performs elastic phase-locked path matching under topological constraints between the current frame's mixed high-frequency sequence and the benchmark high-frequency assembly tolerance sequence. The benchmark high-frequency assembly tolerance sequence is generated based on the benchmark tunnel two-dimensional image frame and its corresponding absolute assembly topological phase coordinates. Specifically, multi-resolution wavelet decomposition is performed on the physical edge geometric space signal of the benchmark tunnel two-dimensional image frame, discarding low-frequency approximation coefficients and reconstructing high-frequency detail coefficients. The obtained high-frequency deviation signals are arranged according to the order of the absolute assembly topological phase coordinates to obtain the benchmark high-frequency assembly tolerance sequence.

[0046] The elastic phase-locked path matching uses a dynamic warping algorithm to calculate the cumulative distance matrix between the two sequences. To avoid excessive distortion, the dominant manifold module 20 introduces a step size constraint on the state transition equation, limiting the continuous step span of the matching path in the horizontal or vertical direction to no more than a set threshold (usually set to 3 to 5 data points). This constrains the spatial monotonicity, local scaling amplitude, and topological order of adjacent tolerance peaks of the matching path. For high-frequency components in the elastic phase-locked path whose span exceeds the scaling limit and cannot satisfy the topological order constraint, or isolated signals that are significantly inconsistent with the confidence of the local evidence field, the dominant manifold module 20 marks them as random vibration components or non-topologically consistent components and reduces their weight in micro-pose locking. The dominant manifold module 20 further calculates the elastic phase-locked residual R based on the cumulative distance per unit length of the elastic phase-locked path. lock R lock It is a non-negative real number. The larger the value, the lower the elastic matching consistency between the current frame's mixed high-frequency sequence and the reference period's high-frequency assembly tolerance sequence.

[0047] Based on the joint index of the current frame, extract multiple historical elastic phase-locked residuals from historical image data, and calculate the mean μ of these multiple historical elastic phase-locked residuals. lock and standard deviation σ lock Thus, the residual judgment threshold T is obtained. lock T lock =μ lock +q lock ·σ lock Among them, q lock This is the anomaly sensitivity coefficient, with a value ranging from 2 to 3, for example, q lock =3. The elastic phase-locked residual R of the current frame... lock With residual judgment threshold T lock Perform a comparison; if R lock >T lock Set the value of the constant Boolean flag for the elastic phase-locked loop residual corresponding to the current frame to 1; if R lock ≤T lock Set the value of the residual abnormal Boolean flag of the elastic phase-locked loop to 0.

[0048] After determining the local neighborhood based on the elastic phase-locked loop path, the dominant manifold module 20 performs one-dimensional normalized cross-correlation matching within this local neighborhood to locally refine the residual microscopic pose bias. The normalized cross-correlation coefficient C during the local refinement process... NCC The calculation formula is: In the formula, K is the total number of sequence sampling points in the local neighborhood; S c (k) represents the discrete sequence value of the mixed high-frequency sequence of the current frame at the k-th sampling point in this local neighborhood, S b (k) represents the discrete sequence value of the high-frequency assembly tolerance sequence at the k-th sampling point within this local neighborhood during the reference period; μ c and μ b These represent the mean of the corresponding local sequences; ε3 is a smoothing coefficient set to prevent the denominator from approaching 0, for example, a value of 10. -6 To avoid computational crashes caused by localized perfect smoothing, the principal manifold module 20 locks the microscopic pose by maximizing the coordinate offset corresponding to the normalized cross-correlation coefficient.

[0049] S203, Deformation Adaptive Manifold Constraint Judgment and Anomaly Isolation Update are Performed. The dominant manifold module 20, based on the absolute assembly topology phase coordinates, camera calibration parameters, and prior design section depth data, performs spatial back projection or geometric constraint calculations on structural primitives such as circumferential joints, longitudinal joints, bolt holes, and cross joints to obtain the current three-dimensional anchor point set. The Euclidean spatial distance between each anchor point in the current three-dimensional anchor point set and the prior design section depth data is calculated and denoted as the spatial deviation. Subsequently, the dominant manifold module 20 performs deformation adaptive manifold constraint judgment and anomaly isolation update based on the spatial deviation and a preset deformation tolerance threshold. The deformation tolerance threshold, used as an engineering control parameter, is preferably set to 0.003 to 0.005 times the tunnel design outer diameter.

[0050] Specifically, when the spatial deviation of the anchor point is less than or equal to a preset deformation tolerance threshold, the dominant manifold module 20 marks the anchor point as a local anchor point, extracts all local anchor points in the current three-dimensional anchor point set to generate a local anchor point residual region, and then uses a non-rigid fitting strategy to perform smooth interpolation and surface reconstruction on the local anchor point residual region to establish a three-dimensional geometric manifold surface that can accurately reflect the actual deformation state of the tunnel. Specifically, the dominant manifold module 20 uses a B-spline surface to absorb the macroscopic low-frequency deformation of the overall tunnel cross-section, and introduces a thin-plate spline energy minimization constraint in the local anchor point residual region to form a local residual compensation field superimposed on the B-spline mainstream shape. The mathematical derivation model of the local residual compensation field is as follows: In the formula, f(x,y) represents the local residual compensation value at the spatial plane coordinates (x,y); a1, a2, and a3 are the affine transformation coefficients that maintain the overall spatial alignment; N is the total number of local anchor points in the local anchor point residual region; w i U(r) represents the weight coefficient of the i-th local anchor point. i ) represents the radial basis function; r i Let (x, y) be the coordinates of the spatial plane and the coordinates of the i-th local anchor point (x, y). i ,y i The actual physical distance between them.

[0051] To ensure stable convergence of the energy minimization constraint in the compensation field, the radial basis function U(r) i The specific definition formula for ) is: U(r) i )=r i 2 ln(r i 2 +ε4). In the formula, ε4 is a set minimum positive real number, for example, a value of 10. -8 This is used to avoid computational anomalies caused by the logarithmic function becoming undefined when the spatial plane coordinates (x, y) coincide with local anchor points (i.e., the radial distance is 0). Simultaneously, it is used to ensure the well-posedness of the thin-plate spline interpolation problem and to solve for the aforementioned weighting coefficients w. iWith the affine transformation coefficients a1, a2, a3, the principal manifold module 20 introduces orthogonal boundary conditions in the solution matrix: , , Furthermore, when establishing the linear equation system matrix to be solved for thin-plate spline interpolation, the principal manifold module 20 uniformly adds a positive regularization parameter ε to the main diagonal of the target kernel matrix. tps This is to prevent the matrix from exhibiting singularities due to the approximate linear arrangement or dense spatial stacking of local anchor points, which would prevent the inversion operation from being completed. ε tps It is a positive real number, and its preferred value range is 10. -8 Up to 10 -4 Finally, the local anchor points that undergo microscopic deformation due to local forces are corrected by the local residual compensation field. The corrected local anchor points are then stored in the manifold memory, thereby updating the manifold control points corresponding to the local anchor points in the manifold memory.

[0052] Conversely, when the spatial deviation of the anchor point exceeds the preset allowable deformation threshold, it indicates that the tunnel may experience structural misalignment or abnormal large deformation. In this case, the dominant manifold module 20 reduces the constraint weight of the prior design section on the current manifold update, shifting the surface reconstruction strategy to a data-driven local manifold fitting mode based primarily on the spatial solution values ​​of the current three-dimensional anchor point set and the continuity of the adjacent topological phase. Similarly, the dominant manifold module 20 writes the anchor point into the macroscopic deformation anomaly flag and stores it in the manifold memory for subsequent cross-dimensional calls by the state control module 10 or the differential attribution module 50.

[0053] Reference Figure 5 In the method provided by this invention, step S300 is specifically implemented as auxiliary-state dual-track estimation and blind zone temporary caching. When the image-based unmanned subway tunnel inspection system switches to auxiliary state due to feature occlusion or severely deteriorated lighting conditions, the auxiliary-state caching module 30 is configured to perform pose estimation and data caching in an environment with missing local features, specifically including the following steps: S301, multiple consecutive sets of tunnel 2D image frames and odometer data are fused for calculation to update the kinematic coordinate orbit. When the system enters the auxiliary state, the auxiliary state cache module 30 acquires multiple consecutive sets of tunnel 2D image frames during the subsequent auxiliary state period, combines them with the current frame to obtain a continuous auxiliary state frame, and uses hardware-triggered timestamps to perform linear interpolation on the time axis for each tunnel 2D image frame in the continuous auxiliary state frame and the odometer data of the wheel odometer, so that the heterogeneous sensor data are aligned under a unified time reference. After alignment, two adjacent tunnel 2D image frames in the continuous auxiliary state frame are recorded as adjacent frames, the dense optical flow field between all adjacent frames is extracted, and the corresponding pixel motion vectors at different scales are estimated by constructing a multi-layer image pyramid, thereby updating the kinematic coordinate orbit.

[0054] Specifically, the auxiliary state buffer module 30 uses the Farnebäck algorithm based on polynomial expansion to perform quadratic polynomial fitting on the neighborhood of each pixel in adjacent frames. By comparing the changes in polynomial coefficients between frames, it calculates the two-dimensional pixel motion vector (Δu, Δv) of each pixel on the image plane. Then, based on the effective pixel ratio of optical flow, the forward and backward optical flow consistency error, and the effectiveness of local texture, it generates the inter-frame optical flow confidence C. flow C flow The value range is [0,1]. Subsequently, the auxiliary state buffer module 30 converts the two-dimensional pixel motion vector (Δu,Δv) into a three-dimensional spatial displacement increment (ΔX,ΔY,Z) with the same dimensions as the wheel odometer. mean ), ΔX=(Δu·Z mean ) / f x ΔY=(Δv·Z) mean ) / f y Where ΔX and ΔY represent the physical displacement increments in the horizontal and vertical directions in the camera coordinate system, respectively, in meters (m); Z mean This represents the average physical depth of the tunnel area corresponding to the two-dimensional image frame of the tunnel, calculated based on prior design cross-sectional depth data, and is expressed in meters (m). x f y These represent the horizontal and vertical focal lengths in the camera intrinsic parameter matrix, respectively, in pixels.

[0055] The above-mentioned three-dimensional spatial displacement increment is denoted as the visual mileage displacement observation vector z. v , i.e. z v =(ΔX,ΔY,Z mean To eliminate erroneous optical flow observations caused by localized reflections, water stains, or dust, the auxiliary state buffer module 30 uses a chi-square test to evaluate the visual mileage displacement observation vector z. v Perform a consistency assessment, calculate the Mahalanobis distance of the observed residuals, and use it as the chi-square test statistic. Let the chi-square test statistic be χ². 2 , χ 2 =(z v -z o ) T (P v +P o ) -1 (z v -z o ), where z o P represents the physical displacement observation vector measured by the wheel odometer. v P represents the visual observation covariance matrix, whose covariance value is positively correlated with the proportion of anomalies in effective optical flow pixels in the optical flow field and the forward and backward optical flow consistency error. oThis represents the observation covariance matrix of the wheeled odometer, which is determined by the noise parameters calibrated at the factory. To avoid visual observations in areas with severely missing features, the observation covariance matrix P... v The presence of rank deficiency or singularity leads to the failure of matrix inversion. The auxiliary state cache module 30, in the visual observation covariance matrix P... v A positive regularization parameter ε5 is added to the main diagonal of the matrix for smoothing. For example, the value of ε5 is set to 10. -6 .

[0056] Furthermore, the auxiliary state caching module 30 will calculate the chi-square test statistic χ². 2 The result is compared with a set confidence threshold, which is determined based on the critical value corresponding to the chi-square distribution table at a 0.95 confidence level. If the chi-square test statistic is less than the critical value, it indicates that the visual observation and the physical odometer observation are consistent. The auxiliary state buffer module 30 uses a Kalman filter framework based on covariance weights to process the visual odometer displacement observation vector z. v and physical displacement observation vector z o Perform weighted fusion to generate the optimal displacement increment X opt X opt =z o +K a ·(z v -z o In the formula, K a K is the Kalman gain matrix. a =P o ·(P o +P v ) -1 If the chi-square test statistic is greater than or equal to this critical value, it indicates that the visual optical flow is in a state of failure or severe interference. The auxiliary state buffer module 30 cuts off the visual observation input, relying only on the physical observation data from the wheeled odometer, i.e., X. opt =z o Finally, the auxiliary state buffer module 30 sequentially accumulates the optimal displacement increment of adjacent frames to the current camera pose, thereby updating the continuous kinematic coordinate trajectory during the auxiliary state.

[0057] S302, based on prior geometric parameters and the extrapolation theory of motion state, predicts the topological phase and maintains the topological predicted coordinate trajectory. While updating the kinematic coordinate trajectory, the auxiliary state buffer module 30 simultaneously constructs the topological predicted coordinate trajectory to provide prior macroscopic geometric constraints. The auxiliary state buffer module 30 extracts the initial topological phase Φ before entering the auxiliary state. t-1 and the standard assembly topology cycle length L of tunnel segments cycle By combining the current linear velocity v and current acceleration a of the inspection vehicle, the theoretical topological prediction phase Φ at the current moment is obtained by extrapolating the acceleration model such as kinematics.t , Φ t =Φ t-1 +(v·Δt+0.5·a·Δt 2 ) / L cycle Where Δt is the time interval between the current frame and the previous valid frame. The initial topology phase is the topology phase at the corresponding moment of the previous valid frame. According to the design data of the subway tunnel to be inspected, the standard assembly topology cycle length L cycle The specific value can be set to 1.2m or 1.5m. The same method is used to smoothly extrapolate the remaining tunnel 2D image frames in the continuous auxiliary state frames to generate the theoretical topology prediction phase at the corresponding time. Through the above smooth extrapolation calculation in the time domain, the auxiliary state buffer module 30 can span occlusion intervals and maintain the spatial continuity of the topology prediction coordinate track.

[0058] S303 utilizes the estimated pose and prior surface to perform a temporary normalized image caching mechanism. To preserve visual recordings during the auxiliary state without disrupting the updated manifold geometric constraints of the primary state, the auxiliary state caching module 30 uses the calculated theoretical topological prediction phase to retrieve the historical manifold control point set for the corresponding position from the manifold memory. Then, combining the dual-track estimated pose provided by the kinematic coordinate track and the topological prediction coordinate track, a preliminary three-dimensional ray intersection unfolding is performed on the currently acquired distorted tunnel 2D image frame. Specifically, the auxiliary state caching module 30 uses the optimal displacement increment provided by the kinematic coordinate track as the microscopic geometric reference and the theoretical topological prediction phase interval determined by the topological prediction coordinate track as the boundary constraint, mapping the optimal displacement increment of the tunnel 2D image frame to the theoretical topological prediction phase interval to generate a comprehensively estimated camera pose.

[0059] Based on this, the auxiliary state buffer module 30 uses the camera projection center determined by the comprehensive calculation of the camera pose as the optical center, and establishes a spatial ray from the optical center through each pixel of the tunnel 2D image frame to the 3D spatial surface. Then, using the rotation and translation matrix obtained from the comprehensive calculation of the camera pose, the ray direction vector in the camera coordinate system is transformed to the globally unified world coordinate system. Subsequently, the 3D coordinates of the intersection point between this spatial ray and the local continuous surface formed by the historical manifold control point set are calculated, and these 3D coordinates are resampled and mapped to a 2D orthogonal plane to generate a temporary normalized image of the tunnel 2D image frame.

[0060] Because the comprehensive estimation of camera pose during the auxiliary state lacks the absolute positional constraints of the real-time multidimensional observation evidence field, there is a risk of cumulative seam misalignment in the macroscopic stitching of the temporary normalized images of each tunnel 2D image frame in consecutive auxiliary state frames. Therefore, the auxiliary state cache module 30 assigns an auxiliary state extrapolation label to the temporary normalized image of each tunnel 2D image frame and writes it into the system's data cache area. Image data with auxiliary state extrapolation labels will not directly participate in the parameter update of the prior reference database. The above data is temporarily retained as an environmental observation record. After the state machine of the image-based unmanned subway tunnel inspection system re-enters the recovered state and completes the global closed-loop geometric correction, the system will then decide whether to trigger post-processing correction for this batch of cached image data based on the closed-loop residual.

[0061] Reference Figure 6 In the method provided by this invention, step S400 is specifically implemented as recovery state verification and reverse pose refeedback. When the system determines that the multidimensional observation evidence field vector of the current frame meets the recovery state feature confidence requirement of the preset observation condition matrix, the recovery refeedback module 40 is configured to perform closed-loop fault tolerance, pose reconstruction, and error truncation. This process specifically includes the following steps: S401, perform topology prediction coordinate trajectory consistency check, calculate the jump deviation between the comprehensively calculated camera pose and the absolute assembly topology phase coordinates. After experiencing feature occlusion or illumination degradation, to avoid erroneous switching of the system state machine caused by occasional environmental noise or isolated similar textures, the state control module 10 executes temporal continuity determination logic. The state control module 10 only confirms the formal entry into the recovery state when the confidence level of the observation evidence field of multiple consecutive tunnel 2D image frames is higher than the preset recovery determination threshold. As a preferred method, the recovery determination threshold is set to 0.85.

[0062] After confirming the transition to the recovery state, the recovery and backfeed module 40 calls the main state manifold module 20 to execute steps S201 and S202 to obtain the absolute assembly topology phase coordinates (X) of the current frame. a ,Y a Z a ) and elastic phase-locked loop residual. Simultaneously, the auxiliary state buffer module 30 extracts the comprehensive calculated camera pose (X) from the most recent auxiliary state output. p ,Y p Z p Then calculate the absolute assembly topology phase coordinates (X). a ,Y a Z a ) and comprehensive estimation of camera pose (X p ,Y p Z p The three-dimensional Euclidean distance between them is defined as the jump deviation D. jump , The jump deviation, in a physical sense, reflects the degree of cumulative drift during the blind zone estimation process.

[0063] To evaluate the reliability of the auxiliary state integrated calculation of camera pose, the recovery and recharge module 40 presets a jump tolerance threshold. This jump tolerance threshold is determined based on the standard design width of a single-ring tunnel segment, and is typically set to 0.1 to 0.2 times the single-ring width. As a preferred approach, for a segment with a standard width of 1.5m, the jump tolerance threshold is specifically set to 0.15m to 0.3m. If the jump deviation is less than or equal to the jump tolerance threshold, proceed to step S402; otherwise, proceed to step S403.

[0064] S402 executes the pose graph optimization refeeding and resampling mechanism to generate normalized images. When the jump deviation is less than or equal to the jump tolerance threshold, it indicates that the auxiliary state comprehensive estimation camera pose is within the cumulative error range of normal occlusion. Under this condition, the recovery refeeding module 40 extracts multiple comprehensive estimation camera poses output in continuous auxiliary states before the system enters the recovery state, defines them as state nodes, and defines the relative pose transformation vectors between adjacent state nodes as constraint edges, thereby constructing the pose graph optimization framework. The global target energy residual is calculated through the objective function E(X) of pose graph optimization. Among them, N e e represents the number of state nodes. i,j Let Ω be the relative pose error vector between state node i and state node j. i,j The information matrix connecting the constraint edges of the two state nodes is mathematically equivalent to the inverse covariance matrix of the edge weights, Ω. i,j =W i,j ·G i,j -1 In the formula, G i,j Let W be the initial relative pose covariance matrix. i,j This is the combined weighting coefficient between state node i and state node j.

[0065] In this embodiment, the inter-frame optical flow confidence C is combined. flow Odometer consistency residual R odom Confidence level C of the observational evidence field evd And the elastic phase-locked residual R of the auxiliary state lock The weighted comprehensive weight coefficient W is generated by weighting. i,j W i,j =α1·C flow +α2·exp(-β1·R odom )+α3·C evd +α4·exp(-β2·R lock), where α1, α2, α3, and α4 are the weighting coefficients for each item, satisfying α1+α2+α3+α4=1. Their specific values ​​are calibrated according to the sensor characteristics, for example, set to 0.3, 0.2, 0.3, and 0.2 respectively. β1 and β2 are positive real constants that adjust the exponential decay rate, with their specific values ​​set between 1.0 and 5.0 (preferably 2.0), used to smoothly map the dimensionless residual to the (0,1] interval. The odometer consistency residual R odom The value is a non-negative real number, obtained by normalizing the Mahalanobis distance or chi-square test statistic between the visual odometer displacement observation vector and the wheel odometer physical displacement observation vector. The elastic phase-locked residual R of the auxiliary state... lock The residual is obtained by interpolating the elastic phase-locked loop residual of the most recent main state before entering the auxiliary state and the elastic phase-locked loop residual obtained when entering the recovery state in chronological order, or by taking the larger of the two as the conservative residual.

[0066] For tunnel 2D image frames with degraded observation conditions or large elastic phase-locked loop residuals, the comprehensive weighting coefficient is reduced, so that in the iterative solution of minimizing the global target energy residual E(X), the absolute drift error is preferentially allocated to the low-confidence segment. To prevent the information matrix from becoming singular or degenerating into a non-positive definite matrix under extremely low confidence, the recovery and recharge module 40 superimposes a positive stability coefficient ε6 (e.g., with a value of 10) on the main diagonal of the information matrix. -5 This ensures the full rank of the graph optimization process for solving the linear equation system. At the solution level, the recovery and backfeeding module 40 uses the Levenberg-Marquardt algorithm to iteratively optimize the aforementioned nonlinear least squares problem. The specific optimization operation is as follows: The recovery module 40 calculates the Jacobian matrix J of the global target energy residual for each state node, which is used to comprehensively infer the camera pose. In each iteration update, according to the formula ΔX=(J... T ΩJ+λI) -1 J T Ωe is used to solve for the pose increment ΔX. In the formula, Ω is the global information matrix; e is the currently calculated error vector; I is the identity matrix; λ is the dynamic damping factor, and the initial value of λ is preferably set to 10. -3 The algorithm iterates up to 1, and then adaptively adjusts based on the decrease in global target energy during the iteration process. When the decrease in error calculated by the iteration is less than a set convergence threshold (e.g., set to 10), the algorithm continues to adjust. -4 When the maximum set number of iterations is reached, preferably 20 to 100, the recovery module 40 determines that the target energy has converged, ends the iteration, and outputs the corrected state nodes.

[0067] After completing the pose graph optimization solution, all the synthesized camera poses during the auxiliary state are corrected and updated. The recovery and refeeding module 40 extracts the temporary normalized image with auxiliary state extrapolation labels from the data buffer. Combining the corrected synthesized camera poses, the corresponding corrected manifold control point set, and the prior design section depth data, the recovery and refeeding module 40 performs a three-dimensional ray intersection operation. After obtaining the three-dimensional spatial intersection points, the recovery and refeeding module 40 uses an equal arc length orthogonal mesh projection method to remap the intersection point coordinates in two-dimensional space. The system divides the orthogonal mesh along the circumferential arc length direction and longitudinal mileage direction of the B-spline physical deformation manifold surface according to a fixed physical resolution (e.g., 1 mm arc length per pixel), and interpolates and resamples the pixel grayscale values ​​of the three-dimensional intersection points to the corresponding nodes of the two-dimensional mesh. Through the resampling process, the system reduces the local image distortion caused by abrupt pose changes and generates a normalized image with suppressed seam misalignment on a macroscopic scale.

[0068] S403 triggers error propagation truncation and reference repositioning logic under extreme conditions. When the calculated jump deviation exceeds the jump tolerance threshold, it indicates that there is long-distance occlusion or continuous loss of structural features during tunnel inspection, and the auxiliary dual-track estimation has deviated beyond the preset tolerance conditions. Under this extreme condition, to prevent severely drifted pose data from damaging the global geometric consistency of the manifold memory, the recovery and refeeding module 40 truncates the error reverse distribution loop of the pose map optimization and triggers the reference repositioning mechanism. Specifically, the recovery and refeeding module 40 does not perform smooth correction on the estimated pose during the auxiliary state, but directly adds a blind zone trajectory loss label to the entire domain of the currently cached temporary normalized image to isolate abnormal data and avoid high-error data from polluting the prior reference database. Subsequently, the system uses the absolute assembly topology phase coordinates obtained in step S401 to cut off the continued propagation of the odometer drift and kinematic extrapolation error accumulated during the auxiliary state to the subsequent state. The recovery module 40 rebinds the current microscopic pose with the absolute assembly topology phase coordinates and restarts the principal manifold constraint logic, so that the subsequently acquired images can be spatially registered based on the re-locked physical reference.

[0069] Reference Figure 7 In the method provided by this invention, step S500 specifically involves constructing a source confidence label matrix and performing artifact differential attribution. In this embodiment, the differential attribution module 50 is configured to establish a source confidence label matrix and perform defect filtering based on causal logic. This process specifically includes the following steps: S501, performs isomorphic mesh mapping and constructs a source confidence label matrix. The differential attribution module 50 uses the tunnel design center axis and the restored B-spline physical deformation manifold surface as the geometric reference to establish an equal-arc-length orthogonal isomorphic mesh coordinate system for the normalized image. Based on this, the differential attribution module 50 constructs a source confidence label matrix pixel-by-pixel with the same scale as this orthogonal isomorphic mesh coordinate system. Each channel of this source confidence label matrix stores different dimensions of system state Boolean flags. Specific Boolean flags include: primary state measurement, secondary state extrapolation, secondary state refeedback correction, macroscopic deformation anomaly, blind zone trajectory loss, manifold control point extrapolation, resampling boundary, illumination degradation, elastic phase-locked loop residual anomaly, local residual compensation enabled, and low-weight refeedback sections of the pose map.

[0070] During the construction process, the differential attribution module 50 assigns values ​​to each pixel node in the current frame based on the output results of the main and auxiliary states before entering the recovery state. For example, for the temporarily normalized image region output by the auxiliary state buffer module 30, its corresponding auxiliary state extrapolation flag is assigned a value of 1, while the remaining normal regions are assigned a value of 0; for pixels located at the edge of image stitching, their corresponding resampling boundary flag is assigned a value of 1. Through this mapping operation, the system solidifies the working state of the multi-source sensors and the processing traces of the underlying algorithm to the pixel-level spatial location, providing direct prior evidence for subsequent artifact removal.

[0071] S502, performs pixel-level isomorphic difference and morphological connected component extraction of the two-phase images, and outputs candidate difference regions. The difference attribution module 50 retrieves the baseline isomorphic tunnel 2D image frame from the prior benchmark database according to the joint index corresponding to the current frame, and performs pixel-by-pixel matrix difference with the normalized image generated in the current frame. Considering the objective existence of attenuation of basic ambient light sources or differences in train illumination between different inspection batches, in order to prevent large-area pseudo-difference noise caused by global illumination changes, before performing matrix difference, the difference attribution module 50 performs histogram specification processing on the normalized image generated in the current frame using the baseline isomorphic tunnel 2D image frame as a reference template, so that the normalized image of the current frame and the baseline isomorphic tunnel 2D image frame are consistent in global grayscale distribution probability. The reference period isomorphic tunnel two-dimensional image frame is generated before the system is first put into the corresponding tunnel section for inspection. Specifically, the system performs three-dimensional ray intersection based on the absolute assembly topology phase coordinates, camera calibration parameters and prior design section depth data of the reference period tunnel two-dimensional image frame, and maps the obtained three-dimensional spatial intersection points to a two-dimensional orthogonal isomorphic grid according to a fixed physical resolution, thereby obtaining the reference period isomorphic tunnel two-dimensional image frame.

[0072] After completing illumination equalization, the current frame normalized image and the baseline isomorphic tunneling 2D image frame are spatially registered to the same orthogonal isomorphic grid coordinate system. Specifically, the differential attribution module 50 calculates the difference or absolute difference of grayscale values ​​at the same pixels between the current frame normalized image and the baseline isomorphic tunneling 2D image frame. Taking the absolute difference ΔI(x,y) as an example, ΔI(x,y)=|I curr (x,y)-I base (x,y)|, where (x,y) are pixel coordinates; I curr (x, y) represents the pixel grayscale value at that coordinate in the current frame's normalized image after illumination equalization; base (x, y) represents the pixel grayscale value of the baseline isomorphic tunnel 2D image frame at this coordinate. Subsequently, a difference matrix is ​​generated based on the absolute grayscale difference between each coordinate in the grid coordinate system. Based on this difference matrix, the normalized image of the current frame is binarized using a set grayscale difference threshold. This grayscale difference threshold is set according to the historical environmental noise level, for example, a grayscale level between 25 and 35.

[0073] Furthermore, the differential attribution module 50 performs morphological connected component analysis on the binarized current frame normalized image. In this embodiment, the differential attribution module 50 uses a two-scan algorithm based on eight-connected neighborhoods to extract connected regions. The specific execution logic is as follows: In the first round of traversing the binarized current frame normalized image pixels, if the current pixel is the target foreground (i.e., binarized to 1), the system checks its left and upper adjacent pixels. If the adjacent pixels have no connected labels, the current pixel is assigned an incremental new label; if the adjacent pixels already have labels, the current pixel is assigned the minimum value among the adjacent labels, and adjacent different labels are recorded as equivalent pairs in the system's equivalence relation table. In the second round of traversing the binarized current frame normalized image pixels, the system uniformly replaces all labels belonging to the same equivalence set with the minimum label value within that set according to the equivalence relation table, thereby completing the division of independent differential candidate regions. After connected component analysis, the differential attribution module 50 extracts several differential candidate regions and calculates the image morphological features of each differential candidate region. The aforementioned image morphological features include the main extension direction vector of the differential candidate region, the gray-level gradient direction vector, and the boundary orientation vector.

[0074] S503 classifies and filters differential candidate regions based on causal logic, attributes and outputs valid surface disease evolution data. Since operations such as pose map refeeding and resampling mapping can easily introduce stretching deformation or splicing artifacts into the image caused by the processing, the differential attribution module 50 calls a rule-based classifier based on geometric orientation consistency to read the image morphological features of each differential candidate region and its corresponding source confidence label matrix.

[0075] Specifically, the differential attribution module 50 extracts the two-dimensional grid coordinates of all foreground pixels within the differential candidate region to calculate the geometric centroid of the differential candidate region, and constructs a two-dimensional covariance matrix of the pixel coordinate group about the centroid. By performing eigenvalue decomposition on this two-dimensional covariance matrix, the system extracts the eigenvector corresponding to the largest eigenvalue, which is used as the principal extension direction vector V of the differential candidate region. ext The main extension direction vector V ext Physically, it represents the macroscopic extension trend of the anomalous region.

[0076] Subsequently, to achieve quantitative evaluation, the system calculates the directional consistency coefficient between the principal extension direction vector of the differential candidate region and the deformation direction vector caused by system operation. The deformation direction vector is the deformation direction vector of the pose re-feedback direction, resampling mapping direction, or manifold extrapolation stitching direction, extracted from the Jacobian matrix or stitching boundary coordinate gradient of the pose reconstruction process. Let the directional consistency coefficient be C. dir C dir =|V ext ·V def | / (|V ext |·|V def |+ε7). Wherein, V ext V def These are the principal extension direction vector and the deformation direction vector, respectively, |V ext |、|V def | represent the main extension direction vectors V and V respectively. ext and deformation direction vector V def The modulus. ε7 is a set minimum positive real bias term (e.g., a value of 10). -8 The technical purpose of this is to avoid calculation anomalies where the denominator of division is zero when the vector magnitude approaches zero. Directional consistency coefficient C dir The value range is [0,1]. The larger the value, the higher the degree of consistency between the main extension direction vector and the deformation direction vector.

[0077] The differential attribution module 50 will use the directional consistency coefficient C dir The value is compared with a set direction determination threshold, typically between 0.85 and 0.90. When the direction consistency coefficient C... dirWhen the direction consistency coefficient is greater than the direction determination threshold, it indicates that the differential candidate region has directional consistency with the deformation direction introduced by the underlying data processing of the system in terms of geometric morphology. At this time, the system checks the source confidence label matrix of the spatial location of the differential candidate region. If the differential candidate region is associated with at least one of the confidence labels such as resampling boundary, macroscopic deformation anomaly, elastic phase-locked loop residual anomaly, or blind zone trajectory loss (i.e., the corresponding Boolean flag is 1), the system determines that the differential candidate region belongs to the system state derived artifact based on the combination of the direction consistency coefficient and the source confidence label. For differential candidate regions determined to be system state derived artifacts, the differential attribution module 50 removes them from the output set of valid surface defect candidate data. Conversely, when the direction consistency coefficient is less than or equal to the direction determination threshold, and the differential candidate region is not associated with the aforementioned confidence labels indicating anomalies or boundaries, it indicates that the grayscale change of the differential candidate region does not meet the attribution conditions of system state derived artifacts. At this time, the system classifies it as valid surface defect evolution data and outputs it. After multi-dimensional cross-filtering based on the above causal logic, the differential attribution module 50 finally outputs effective surface disease evolution data for subsequent system execution of disease quantification and maintenance decisions. Example 3

[0078] The present invention also provides a computer device, comprising: a processor and a memory. The processor is used to execute a computer program for implementing an image-based unmanned inspection method for subway tunnels. The memory is used to store the computer program.

[0079] The memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk. Example 4

[0080] To further illustrate the operating mechanism and technical effects of the present invention in a real engineering environment, this embodiment is explained in conjunction with specific practical application cases and experimental verification data.

[0081] Taking a routine nighttime inspection of a shield tunnel section on a major subway line in China as an example, the inspection trolley travels at a constant speed, carrying the image-based unmanned subway tunnel inspection system of this invention to perform full-section data acquisition. This inspection section faces typical complex engineering conditions: from K14+200 to K14+250, the track bed surface is slippery due to long-term localized water seepage; subsequently, from K14+230 to K14+280, a short-term overheating protection power failure of the main strobe lighting unit on the inspection trolley, coupled with dense water mist from strong convection at the ventilation shaft opening, resulted in severe underexposure of the camera and complete obscuring of the tunnel segment surface texture, forming a continuous visual blind spot approximately 50 meters in length. Prior design data showed no obvious defects in this section during the baseline period, but before this inspection, a real structural spalling occurred at K14+265 due to minor external soil movement.

[0082] When the trolley enters the K14+200 section, the system initially enters the main state. The state control module 10 simultaneously captures image frames and high-frequency wheel odometer data. At this time, the wet track bed causes the odometer to instantaneously acquire reverse pulses and forward slip signals. The system immediately generates a dynamic odometer search domain based on the calibration parameters, effectively encapsulating the nonlinear slip error. Within this dynamic odometer search domain, the main state manifold module 20 stably extracts structural primitives such as circumferential seams and bolt holes. Through elastic path matching of the high-frequency tolerance sequence, it completely absorbs the microscopic pose bias and updates the absolute assembly topology phase and manifold control points written to the current frame.

[0083] As the trolley enters the K14+230 blind zone, the state control module 10 calculates and finds that the significance of the grayscale projection valleys and the local texture information entropy of multiple consecutive two-dimensional image frames drop sharply. The multidimensional observation evidence field vector is far below the confidence requirement of the main state feature, and the system is forced to transition to the auxiliary state.

[0084] Reference Figure 8 In the early stages of the auxiliary state ( Figure 8 In the 100-150 frame range, due to the faint residual reflection on the surface of the tube segment, the auxiliary buffer module 30 fuses dense optical flow and odometer displacement increments to maintain the kinematic coordinate track; when it completely enters the core blind zone, the chi-square test determines that visual observation has failed, and the system decisively cuts off the optical flow input. The dashed line in the figure is the trajectory calculated by the traditional pure odometer. Due to the failure to eliminate slippage in the early stage, it diverges rapidly in the area without visual constraints, with a cumulative longitudinal drift of more than 0.8 meters; the solid line in the figure is the topology prediction coordinate track maintained by this invention, which uses the prior geometric parameters and kinematic state before entering the blind zone to perform smooth extrapolation, forcibly binding the trajectory deviation within the theoretical tolerance zone ( Figure 8 Within the gray shaded area. During this period, the system retrieves the historical manifold control point set from the manifold memory to generate a temporary normalized image, and injects all of them into the auxiliary state extrapolation label and temporarily stores them in the cache to avoid polluting the prior baseline database.

[0085] When the trolley leaves K14+280 and reaches the normal section, the supplementary lighting group restarts and restores, and environmental features reappear. The state control module 10 extracts high-confidence structural primitives within 5 consecutive frames and instructs the system to transition to the recovery state. The recovery and recharge module 40 then acquires the currently relocked actual absolute assembly topology phase and verifies it against the coordinates calculated from the topology prediction coordinate track at the end of the auxiliary state. The measured jump deviation is 0.18 meters, within the set jump tolerance threshold. The system then initiates the pose graph optimization framework, using the absolute physical coordinates at both ends of the blind zone as rigid anchor points. The 0.18-meter absolute drift error is then smoothly distributed in reverse according to the covariance weight to all implicit pose nodes during the auxiliary state, completing the process. Figure 8 Closed-loop re-irrigation correction of the solid line trajectory is performed, and resampling is carried out according to the orthogonal grid to generate a seamlessly connected normalized image.

[0086] After the entire acquisition is completed, the differential attribution module 50 intervenes to perform artifact filtering. Because the pose map optimization introduces reverse stretching deformation in the blind zone estimation section to forcibly stitch together errors, if pixel-level differential is performed directly, a large number of patches caused by texture misalignment will inevitably be generated at the image stitching point.

[0087] Reference Figure 9 After completing histogram specification and matrix differencing, the difference attribution module 50 extracts numerous difference candidate regions. Figure 9 Scattered points). Comparison revealed that a large number of scattered points were densely distributed in regions with x-coordinates greater than 0.85. Furthermore, the system retrieved the spatial locations of these regions and found that their source confidence label matrices were automatically written with auxiliary state refeedback correction or resampling boundary flags. The rule-based classifier, based on geometric orientation consistency, uses causal logic to classify these highly directionally consistent difference regions (…). Figure 9 Hollow dots were identified as system state-derived artifacts and were directly removed.

[0088] Conversely, the differential candidate region formed by the actual structural exfoliation disease at K14+265 exhibits an irregular shape in its main extension direction, with a calculated directional consistency coefficient of only 0.32, and severe abrupt changes in local gray-scale gradients. Figure 9 (Marked with an asterisk). This area successfully avoided the artifact filtering logic, did not match any abnormal derived tags, and was ultimately accurately extracted by the system and output as valid surface disease evolution data.

[0089] To verify the overall engineering effectiveness of the above technical solution, a comparative test was conducted based on a real historical inspection dataset of shield tunnels. The dataset centrally reproduced the visual blind spot conditions caused by hardware failure or extreme water mist in the ventilation shaft.

[0090] Regarding spatial reference continuity, conventional visual SLAM schemes without manifold geometric constraints are prone to decimeter-level jump deviations when exiting blind zones, leading to severe macroscopic tearing in the stitching of preceding and following images. In contrast, the scheme of this invention, through auxiliary-state dual-track extrapolation and a recovering-state error refeedback mechanism, significantly reduces the jump deviation when exiting blind zones. After image-optimized refeedback and resampling, the macroscopic seam misalignment of the normalized image is drastically reduced from tens of pixels in traditional schemes to sub-pixel or even a very small number of pixels.

[0091] Regarding the robustness of defect detection, traditional image direct difference methods generate a large number of false alarms for suspected defects under such conditions, resulting in an extremely high false alarm rate. The solution of this invention relies on an underlying confidence matrix and a morphological causal attribution mechanism to significantly reduce invalid false alarms while ensuring a high recall rate for genuine defects, thereby improving the artifact filtering rate.

[0092] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An image-based unmanned inspection method for subway tunnels, characterized in that, Includes the following steps: Simultaneously acquire tunnel two-dimensional image frames and mileage data, extract multiple image feature parameters in the dynamic mileage search domain to construct a multidimensional observation evidence field vector. The image feature parameters include grayscale projection valley significance coefficient, local texture information entropy, anisotropic response parameter and unidirectional phase consistency parameter. Compare the multidimensional observation evidence field vector with the preset observation condition matrix, and perform state transition judgment in combination with the current system state to drive the state machine to transition between the main state, auxiliary state and recovery state. The observation condition matrix is ​​composed of confidence requirement vectors of the main state, auxiliary state and recovery state. In the main state, the absolute assembly topology phase coordinates are matched, the mixed high-frequency sequence is extracted to perform elastic phase-locked path matching and micro pose locking, and the manifold control point update and anomaly isolation update are performed. When in auxiliary state, the optimal displacement increment is calculated to update the kinematic coordinate orbit, the theoretical topology prediction phase is extrapolated to maintain the topology prediction coordinate orbit, and the kinematic coordinate orbit and the topology prediction coordinate orbit are combined to generate a comprehensive calculated camera pose. The comprehensive calculated camera pose and the historical manifold control point set are combined to generate a temporary normalized image. In the recovery state, a topology prediction coordinate trajectory consistency check is performed to generate a jump deviation. Based on the jump deviation, pose error optimization and refeedback are performed, along with mesh unfolding and resampling. This generates a normalized image or triggers reference repositioning by truncating the optimization loop. The pose error optimization and backfeeding is as follows: extract all the comprehensive inference camera poses during the auxiliary state, define each comprehensive inference camera pose as a state node, define the relative pose transformation vector between adjacent state nodes as a constraint edge, construct a pose graph optimization framework by combining state nodes and constraint edges, perform pose graph optimization based on the pose graph optimization framework, and correct the comprehensive inference camera pose. Establish a source confidence label matrix for standardized images and extract differential candidate regions. Combine the differential candidate regions with the source confidence label matrix to perform attribution determination, filter system state-derived artifacts, and output effective surface disease evolution data.

2. The image-based unmanned inspection method for subway tunnels according to claim 1, characterized in that, Matching absolute assembly topology phase coordinates includes: Using edge detection operators, Hough linear transform and Hough circular transform, the intersection points of annular joints, longitudinal joints, bolt holes and cross joints are extracted from the tunnel 2D image frame as structural primitives and defined as nodes. The relative spatial distance and angular relationship between the extracted nodes are defined as edges, and a directed observation topology segment is established. The differences in node type, side length Euclidean distance, and included angle cosine value between the directed observation topology fragment and the global design diagram structure in the prior benchmark database are calculated. Independent weights are assigned to the differences in node type, side length Euclidean distance, and included angle cosine value, and a weighted sum is performed to generate a comprehensive topology similarity score. The position coordinates with the highest comprehensive topology similarity score that are greater than the set judgment threshold are set as the absolute assembly topology phase coordinates.

3. The image-based unmanned inspection method for subway tunnels according to claim 1, characterized in that, Extracting mixed high-frequency sequences and performing elastic phase-locked path matching and micro-pose locking includes: Multi-resolution wavelet decomposition is performed on the geometric spatial signal of the physical edge of the tunnel two-dimensional image frame. Low-frequency approximation coefficients are discarded and high-frequency detail coefficients are reconstructed. High-frequency deviation signals reflecting the small differences of the tunnel segments are extracted to generate a hybrid high-frequency sequence. Based on the multidimensional observation evidence field vector, absolute assembly topology phase coordinates, and historical topology prediction deviation, a soft constraint weight window is generated. Under the guidance of the soft constraint weight window, a dynamic warping algorithm is used to calculate the cumulative distance matrix between the mixed high-frequency sequence and the benchmark period high-frequency assembly tolerance sequence. The elastic phase-locked path is extracted by introducing step size constraint, and the elastic phase-locked residual is calculated based on the elastic phase-locked path. Within the local neighborhood determined by the elastic phase-locked path, a one-dimensional normalized cross-correlation matching is performed on the hybrid high-frequency sequence and the reference period high-frequency assembly tolerance sequence to complete the locking of the micro pose.

4. The image-based unmanned inspection method for subway tunnels according to claim 2, characterized in that, Perform manifold control point updates and anomaly isolation updates, including: Spatial back projection is performed on the intersection of circumferential joints, longitudinal joints, bolt holes and cross joints to obtain the current three-dimensional anchor point set. The Euclidean spatial distance between each anchor point in the current three-dimensional anchor point set and the prior design section depth data is calculated, and the Euclidean spatial distance is used as the spatial deviation of each anchor point. Extract all anchor points with spatial deviations less than or equal to the preset deformation tolerance threshold, generate local anchor point residual regions, use B-spline surfaces to absorb the macroscopic low-frequency deformation of the overall tunnel cross section, introduce thin plate spline energy minimization constraints in local anchor point residual regions, form a local residual compensation field superimposed on the B-spline surface, and update the corresponding manifold control points in the manifold memory bank according to the local residual compensation field. Extract anchor points whose spatial deviation is greater than the preset deformation tolerance threshold, and write the anchor points into the macroscopic deformation anomaly flag in the manifold memory.

5. The image-based unmanned inspection method for subway tunnels according to claim 4, characterized in that, The optimal displacement increment is calculated to update the kinematic coordinate trajectory. The theoretical topological prediction phase is extrapolated to maintain the topological predicted coordinate trajectory. The kinematic coordinate trajectory and the topological predicted coordinate trajectory are combined to generate a comprehensively calculated camera pose. Finally, the comprehensively calculated camera pose and historical manifold control point sets are combined to generate a temporary normalized image, including: Extract multiple sets of tunnel 2D image frames acquired during the system's auxiliary state, record two adjacent sets of tunnel 2D image frames as adjacent frames, calculate the polynomial expansion coefficients of adjacent frames, compare the changes in the polynomial coefficients of adjacent frames to obtain 2D pixel motion vectors, combine the 2D pixel motion vectors with the camera intrinsic parameter matrix to generate 3D spatial displacement increments, and record the 3D spatial displacement increments as visual mileage displacement observation vectors. The observation residual between the visual odometer displacement observation vector and the wheel odometer physical displacement observation vector is calculated using the chi-square test. The visual odometer displacement observation vector and the wheel odometer physical displacement observation vector are weighted and fused using a Kalman filter framework to output the optimal displacement increment. The kinematic coordinate trajectory is updated based on the optimal displacement increment of all adjacent frames and the current camera pose. Extract the initial topological phase before entering the auxiliary state, combine the standard assembly topological cycle length of the tunnel segment, the current linear velocity and current acceleration of the vehicle, calculate the theoretical topological prediction phase of each tunnel two-dimensional image frame in the continuous auxiliary state frame based on the uniform acceleration kinematic model, and generate the topological prediction coordinate trajectory by combining the theoretical topological prediction phases of all tunnel two-dimensional image frames. Based on the theoretical topological prediction phase of the tunnel 2D image frame, the historical manifold control point set corresponding to the location is retrieved from the manifold memory. The optimal displacement increment of the tunnel 2D image frame is mapped to the theoretical topological prediction phase interval determined by the topological prediction coordinate track to generate the corresponding comprehensive estimated camera pose. The temporary normalized image is generated by combining the comprehensive estimated camera pose with the historical manifold control point set.

6. The image-based unmanned inspection method for subway tunnels according to claim 1, characterized in that, Establish a source confidence label matrix for normalized images and extract differential candidate regions, including: Based on the orthogonal isomorphic grid coordinate system, a source confidence label matrix containing Boolean flags of the system state is constructed pixel by pixel; Using the baseline isomorphic tunnel two-dimensional image frame as a reference template, the normalized image is subjected to histogram specification processing; In the orthogonal isomorphic grid coordinate system, the gray-level absolute difference matrix between the reference period isomorphic tunnel two-dimensional image frame and the normalized image after histogram specification is calculated, and the gray-level absolute difference matrix is ​​binarized by setting the gray-level difference threshold. The two-scan algorithm based on eight-connected neighborhoods extracts independent connected regions from the binary gray-level absolute difference matrix and outputs differential candidate regions.

7. The image-based unmanned inspection method for subway tunnels according to claim 1, characterized in that, Attribution determination is performed by combining differential candidate regions and source confidence label matrices, filtering system state-derived artifacts, and outputting valid surface defect evolution data, including: Extract the grid coordinates of all foreground pixels within the differential candidate region to construct a two-dimensional covariance matrix. Obtain the eigenvector corresponding to the largest eigenvalue through eigenvalue decomposition as the main extension direction vector. Extract the deformation direction vector of the system operation and calculate the directional consistency coefficient between the main extension direction vector and the deformation direction vector. If the directional consistency coefficient of the differential candidate region is greater than the directional determination threshold, and the corresponding source confidence label matrix contains a resampling boundary flag, a macroscopic deformation anomaly flag, an elastic phase-locked loop residual anomaly flag, or a blind zone trajectory loss flag, then the differential candidate region is determined to be a system state-derived artifact and is removed; otherwise, the differential candidate region is determined to be valid surface disease evolution data and is output.

8. A system for implementing the image-based unmanned inspection method for subway tunnels as described in any one of claims 1-7, characterized in that, include: The state control module is configured to synchronously acquire tunnel two-dimensional image frames and mileage data, generate a dynamic mileage search domain and output a multi-dimensional observation evidence field vector, and drive the state machine to transition between the main state, auxiliary state and recovery state. The dominant manifold module is configured to match the absolute assembly topology phase coordinates, extract the mixed high-frequency sequence to perform microscopic pose locking, and update the manifold control points when it is in the dominant state. The auxiliary state cache module is configured to, when in auxiliary state, calculate the optimal displacement increment to update the kinematic coordinate trajectory, extrapolate the theoretical topology prediction phase to maintain the topology prediction coordinate trajectory, and generate a temporary normalized image and store it in the data cache area. When the recovery module is configured to be in recovery mode, it calculates the jump deviation and generates a normalized image or triggers a reference repositioning based on the jump deviation. The differential attribution module is configured to generate a source confidence label matrix and differential candidate regions for normalized images. It combines the source confidence label matrix and differential candidate regions to generate directional consistency coefficients, eliminates system state-derived artifacts based on the directional consistency coefficients, and outputs effective surface disease evolution data.

Images