A risk-driven large linear site air-ground collaborative hierarchical inspection method

Abstract

This invention discloses a risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites. The method includes aerial acquisition of imagery and point cloud data of the linear archaeological site to be inspected, generating a digital surface model; calculating the risk value of each cell in the digital surface model based on at least one of topographic, structural, material, and environmental indicators; determining ground detection candidate areas based on the risk values; then performing accessibility scoring on the ground detection candidate areas and generating a detection task sequence based on the scoring results; and having a ground detection device arrive at the points sequentially according to the detection task sequence. Through pose-synchronized back-projection voxelization and fusion of multi-source surface features, a three-dimensional convolutional neural network is driven to achieve accurate detection of internal cavities within the archaeological site. This application overcomes the problems of fragmented surface observation and internal detection, inefficient resource scheduling, and high maintenance costs of three-dimensional models in traditional inspections, thus forming a scalable, reusable, and sustainably updated digital twin inspection framework for archaeological sites.

Description

Technical Field

[0001] This invention relates to the field of digital preservation of cultural relics and intelligent robot technology, and more specifically to a risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites. In particular, it relates to a method that combines wide-area aerial perception, risk-driven scheduling, ground contact detection, and voxelized digital twins for air-ground collaborative hierarchical inspection of large-scale linear archaeological sites. Background Technology

[0002] Large linear archaeological sites (such as the Great Wall) are characterized by their large scale, wide spatial extension, complex terrain, and diverse types of construction materials. Long-term exposure to weathering, rain, freeze-thaw cycles, and biological erosion makes them prone to both surface deterioration and internal structural defects. For the routine inspection and safety assessment of these sites, traditional methods relying on manual on-site inspections are not only labor-intensive and inefficient, but also pose certain safety risks in complex terrain conditions, making them unsuitable for the practical needs of large-scale, routine inspections.

[0003] In existing technologies, UAV aerial surveying or laser scanning can acquire surface images and point cloud data of archaeological sites in a short time and construct 3D surface models, showing significant advantages in large-scale coverage. However, these methods mainly reflect the external appearance of the site and are difficult to directly obtain the true state of the interior of the bricks and stones, the deep layers of the structure, or hidden parts. Furthermore, they are prone to creating voids or low-coverage areas in the model due to steep slopes, vegetation obstruction, or limited viewing angles. When voids or incomplete data appear in the aerial survey 3D model, manual model patching is still the primary method, which is difficult to adapt to the large-scale data processing requirements of long-term, repetitive inspections. In contrast, ground-based contact non-destructive testing technology can detect the internal structure of archaeological sites, but its operating range is limited and its coverage efficiency is low.

[0004] However, the lack of high-precision spatial registration between internal detection and aerial 3D models makes it difficult to uniformly express and comprehensively analyze the detection results on a global scale, which can easily lead to a disconnect between surface data and internal detection data.

[0005] Therefore, how to achieve air-ground collaboration and improve the efficiency and accuracy of inspections of large linear archaeological sites is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] To address the problems of low efficiency and high safety risks in existing large-scale linear cultural site inspections, difficulty in obtaining internal structural information from aerial inspections, limited coverage of ground contact inspections and lack of accurate registration with aerial data, as well as the tendency for voids in aerial survey 3D models and reliance on manual model repair, this application proposes a risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear sites to overcome or at least partially solve these problems.

[0007] To achieve the above objectives, embodiments of the present invention provide a risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites, comprising:

[0008] Aerial images and point cloud data of linear archaeological sites to be inspected are collected to generate digital surface models; Based on at least one of terrain, structural, material and environmental indicators, calculate the risk value of each cell in the digital surface model. Candidate areas for ground detection are determined based on the risk values; The ground detection candidate areas are scored for accessibility, and sorted according to the scoring results to generate a detection task sequence; The ground detection device arrives at the designated locations sequentially according to the detection task sequence. Through pose-synchronized back-projection voxelization and fusion of multi-source surface features, it drives a three-dimensional convolutional neural network to achieve accurate detection of cavities inside the site.

[0009] In one optional embodiment, the topographic index is a local slope value, a local curvature value, or a weighted value of both; the structural index is the crack density in the texture anomaly region; the material index is a weighted value of the normalized degradation index calculated based on the image and the proportion of the reflectance intensity anomaly area determined based on the point cloud data; and the environmental index is a weighted value of the normalized water index and the normalized vegetation index.

[0010] In one optional embodiment, the risk value of each cell in the digital surface model is calculated according to the following formula:

[0011] In the formula, Indicates the first The risk value of each unit. This represents the risk calculation function. Indicates topographical indicators, Indicates structural indicators, Indicates material specifications, Indicates environmental indicators.

[0012] In one optional embodiment, reachingability scoring of the ground detection candidate region includes:

[0013] In the formula, Indicates the risk value. The weights representing risk values Indicates accessibility score, This indicates the weight of the accessibility score. Indicates the detection value score. This indicates the weight of the detection value score.

[0014] In one optional embodiment, the accessibility score is obtained through the following steps: The path cost is determined as follows;

[0015] in, Represents path cost, This represents the total path length from the ground detection device to the current point in the ground detection candidate area. This represents the average slope cost along the path. This represents the cost of obstacle density in the path. This indicates the estimated energy consumption along the path. These are the fusion weights for total path length, average slope cost, obstacle density cost, and energy consumption, respectively. A reachability score is determined based on the path cost;

[0016] In the formula, Accessibility is scored.

[0017] In one optional embodiment, the detection value score calculation process is as follows:

[0018] In the formula, Indicates the defect confidence factor. The weights corresponding to the defect confidence factors. Indicates structural importance factor, The weights corresponding to the structural importance factors are... Indicates the inspection timeliness factor. The weights corresponding to the inspection timeliness factor.

[0019] In one optional embodiment, when the ground detection device arrives at the points sequentially according to the detection task sequence, the posture of the robotic arm probe in the ground detection device is adjusted according to the surface normal vector. The steps include: Obtain the normal vector of the surface at the target point; The direction opposite to the normal vector is taken as the axis vector of the robotic arm probe; The included angle error and rotation axis are determined based on the probe axis vector and the desired direction vector of the robotic arm, respectively. An attitude increment rotation matrix is ​​generated based on the angle error and the rotation axis, which is used to iteratively correct the pose of the robotic arm end effector until the angle between the probe axis and the surface normal is less than a preset threshold.

[0020] In an optional embodiment, when the ground detection device arrives at the points sequentially according to the detection task sequence, it performs impedance force control, wherein the contact pressure is:

[0021] In the formula, For contact force, For the desired position, For the desired speed, For actual location, For actual speed, 、 These are control parameters used to maintain a constant contact pressure and reduce mechanical stress on the site.

[0022] In one optional embodiment, a three-dimensional convolutional neural network is driven to accurately detect internal cavities in a site by fusion of pose-synchronized back-projection voxelization and multi-source surface features; including: S1. Collect probe echo data through the robotic arm probe in the ground detection device, and record the probe pose and timestamp; S2. Based on the pose information and the depth information of the echo data, the internal response is projected onto a preset three-dimensional voxel grid, and the probability of the defects penetrating the voxel is accumulated and calculated using a probability update algorithm to generate an initial defect probability voxel map. S3. Spatially align and feature-stitch the digital surface model and the initial defect probability voxel map in a multi-scale voxel grid to construct a three-dimensional digital twin model containing surface texture, geometric information, material density estimation and defect probability. S4. Divide the fused 3D digital twin model into voxel blocks and input them into a pre-trained 3D convolutional neural network model to output a voxel-level hole probability distribution; perform thresholding on the hole probability distribution and use a clustering algorithm to merge connected components to generate hole detection results that include location, size, confidence, and priority.

[0023] In one optional embodiment, when the confidence level in the detection result is lower than a preset quality threshold or a valid void is detected, a supplementary testing instruction or a go-around mission plan is automatically generated; and the generated void detection results and repair markings are sent to the maintenance work order. Simultaneously, all raw images, point cloud data, probe echo data, probe pose data, and voxelized model data are stored with unified geographic coordinates and associated with timestamps to support long-term degradation trend monitoring.

[0024] This invention provides a risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites. The method constructs a surface model of the site and assesses structural risks through aerial inspection. Risk-driven air-ground collaborative inspection tasks are generated, and ground devices perform flexible contact detection on high-risk areas. Finally, multi-source data are fused to construct a three-dimensional digital twin model and achieve automatic identification and labeling of voids.

[0025] Compared with existing technologies, the beneficial effects include at least the following: First, by adopting a risk-driven air-ground collaborative hierarchical inspection strategy, ground detection resources are prioritized for high-risk areas, significantly improving inspection efficiency and targeting. Second, by back-projecting the internal detection data based on the probe pose and fusing it with voxelized multi-source data, a digital twin representation of the site that is "integrated inside and out" can be achieved, which facilitates quantitative assessment of the structural state and restoration decisions. Third, by introducing an automatic detection and annotation mechanism for three-dimensional voids, the automatic identification of voids in aerial survey three-dimensional models and the generation of model repair tasks can be realized, reducing the workload of manual model repair and improving the integrity of the model. Fourth, by using flexible contact and force control strategies to perform ground inspections, the potential mechanical damage to the site during the inspection process can be effectively reduced, thereby improving the safety of cultural relic protection. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0027] Figure 1 This is a flowchart of a risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites, provided in an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of 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.

[0029] This invention discloses a risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites. This method organically combines rapid aerial modeling, risk-assessment-based inspection task scheduling, precise ground-based contact detection, and voxelized multi-source data fusion with automatic cavity detection. It achieves air-ground collaboration, hierarchical inspection, and integrated digital representation, thereby improving the efficiency, accuracy, and long-term management capabilities of large-scale linear archaeological site inspections. Specific implementation methods are described below.

[0030] This invention provides a risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites, such as... Figure 1 As shown, it includes: Aerial images and point cloud data of linear archaeological sites to be inspected are collected to generate digital surface models; Based on at least one of terrain, structural, material and environmental indicators, calculate the risk value of each cell in the digital surface model. Candidate areas for ground detection are determined based on the risk values; The ground detection candidate areas are scored for accessibility, and sorted according to the scoring results to generate a detection task sequence; The ground detection device arrives at the designated locations sequentially according to the detection task sequence. Through pose-synchronized back-projection voxelization and fusion of multi-source surface features, it drives a three-dimensional convolutional neural network to achieve accurate detection of cavities inside the site.

[0031] In a preferred embodiment, aerial data acquisition is achieved using drones, and the ground detection device is configured as a ground flexible contact detection robot.

[0032] The first step involves the UAV collecting visible light imagery, multispectral imagery, and sparse point cloud data with over 70% overlap along a planned flight path. SFM / MVS and LiDAR-visual registration are then performed from the cloud / ground terminals to generate dense point cloud and orthophotos, resulting in a digital surface model (DSM). Subsequently, the point cloud undergoes preprocessing, including statistical outlier removal and downsampling of nearest-distance redundant points, to improve point cloud quality and preliminarily label point cloud hole masks.

[0033] The second step is to conduct a risk assessment on the digital surface model, calculate the risk value of each cell in the model, obtain a spatialized risk map, and classify it according to a threshold; further, select the cells corresponding to the top 10-20% of the R values ​​in the high-risk level as ground detection candidate areas.

[0034] In this embodiment, the risk assessment indicators include topographic indicators, structural indicators, material indicators, and environmental indicators.

[0035] In some implementations, the terrain index is a local slope value, a local curvature value, or a weighted average of both. Specifically, the local slope value is obtained by calculating the elevation gradient of adjacent cells, and is expressed as:

[0036] in, The grid cell elevation value. , These represent the planar coordinates in space. and Let represent the first-order partial derivatives of elevation along the x and y directions, respectively. Assuming the current cell position is (i, j), then:

[0037] and These are the elevation values ​​of the current cell to its left and right adjacent cells in the x-direction. This represents the grid spacing.

[0038] Furthermore, second-order partial derivatives are calculated for the elevation of the same cell and its neighborhood, and the Laplace operator is used to characterize the local curvature change; expressed as:

[0039] in, It is the Laplace operator.

[0040] In this embodiment, the local slope value and the local curvature value are normalized respectively, and the maximum value or weighted value of the two is taken as the topographic index S of the grid cell, thereby reflecting the steepness of the surface of the site and the degree of structural change.

[0041] In some implementations, the structural index is the crack density in texture anomaly regions. Specifically, firstly, a multi-directional Gabor filter is applied to the orthophoto image to extract texture responses at different scales and orientations, obtaining candidate texture anomaly regions; then, Canny edge detection is performed on the image to extract crack edge lines, and simultaneously, the ratio of the total length of crack lines in each cell to the area of ​​that cell is calculated to obtain the crack density index, expressed as:

[0042] In the formula, The sum of the lengths of the line elements in the crack within the lattice cell. The area of ​​a grid cell.

[0043] For texture anomaly regions, the abnormal response intensity, the area ratio of the abnormal region, and the crack connectivity length can be further statistically analyzed and normalized to obtain the final texture anomaly / crack density index C, which reflects the degree of surface damage and crack development.

[0044] In some implementation schemes, the material index is a weighted value of a normalized degradation index calculated based on the imagery and the proportion of anomaly areas in reflectance intensity determined based on the point cloud data. This includes first calculating the normalized degradation index based on multispectral imagery to characterize the degree of fading, weathering, and aging of the surface material; wherein the degradation index can be composed of reflectance characteristics from multiple bands and normalized. Further, the proportion of areas with anomaly in reflectance intensity in the point cloud is combined to supplement the characterization of local material density changes, surface erosion, and the risk of hollowing. Finally, the multispectral degradation index and the proportion of anomaly areas in reflectance intensity are weighted and fused according to preset weights to obtain the material index. This index is used to reflect the weathering level and structural deterioration degree of the archaeological site's construction materials.

[0045] In some implementation schemes, the environmental indicator is a weighted value of the Normalized Difference Water Index (NDWI) and the Normalized Difference Vegetation Index (NDVI). This includes calculating the NDWI based on multispectral or remote sensing imagery to characterize surface humidity, water accumulation, or water content anomalies; and then calculating the NDVI to characterize vegetation cover and its occlusion impact on inspections. The NDWI and NDVI are then linearly weighted and fused to form the environmental indicator, with higher weights given to areas with higher humidity, denser vegetation cover, or water accumulation to reflect their interference with UAV observation, ground access, and subsequent detection. A higher environmental indicator indicates a more significant environmental impact and a higher inspection difficulty in that area.

[0046] Finally, based on the above indicators, the risk value of each cell in the digital surface model is calculated using the following formula:

[0047] In the formula, Indicates the first The risk value of each unit. The risk calculation function is represented by a linear weighted model, which is preferred. Indicates topographical indicators, Indicates structural indicators, Indicates material specifications, Indicates environmental indicators.

[0048] The third step is to score the accessibility of each point in the ground detection candidate area, sort them according to the scoring results, generate a detection task sequence, send the detection task sequence to the ground detection device, and plan the optimal path. If communication is interrupted or the signal is weak, a retreat path and retreat strategy are formulated. This application adopts a hierarchical inspection architecture of air-ground cooperation, and drives the ground robot to operate precisely through the intelligent screening of UAVs.

[0049] In this embodiment, based on accessibility scoring and detection value score The comprehensive accessibility score for the ground detection candidate area is calculated using the following formula:

[0050] In the formula, Indicates the risk value. The weights representing risk values This represents the weight of the accessibility score. This indicates the weight of the detection value score.

[0051] The accessibility score is calculated by comprehensively considering the path length of candidate detection points, terrain navigability, obstacle density, and energy consumption. Specifically, the optimal path from the current location to candidate detection points is first constructed based on the digital surface model of the site, and the total length of this path is then calculated. Average slope cost Obstacle density cost and projected energy consumption Among them, the slope cost can be represented by the average or maximum slope of each segment of the path, the obstacle density cost can be obtained by statistically analyzing the number of gravel, steps, vegetation or fracture structures per unit length along the path, and the energy consumption can be estimated by the robot kinematics model or empirical energy consumption model.

[0052] The path cost is then obtained by normalizing each cost term and performing a weighted sum. :

[0053] In the formula, Represents path cost, This represents the total path length from the ground detection device to the current point in the ground detection candidate area. This represents the average slope cost along the path. This represents the cost of obstacle density in the path. This indicates the estimated energy consumption along the path. These are the fusion weights for total path length, average slope cost, obstacle density cost, and energy consumption, respectively. Furthermore, the path cost is mapped to an reachability score;

[0054] In the formula, For accessibility scoring, The larger the value, the easier it is to reach the candidate detection point and the lower the inspection execution cost; The smaller the value, the more complex the terrain, the more difficult the passage, or the higher the energy consumption at that point.

[0055] In this embodiment, the detection value score is obtained by normalizing and weighting the defect confidence factor, structural importance factor, and inspection timeliness factor for the region where the candidate detection point is located, and then fusing them together. :

[0056] In the formula, Indicates the defect confidence factor. The weights corresponding to the defect confidence factors. Indicates structural importance factor, The weights corresponding to the structural importance factors are... Indicates the inspection timeliness factor. for The inspection timeliness factor corresponds to a weight. In this embodiment, The higher the value, the more worthy the candidate detection point is of detailed ground detection, meaning the greater its contribution to improving the understanding of the site's structure, the identification of defects, and the decision-making process for restoration.

[0057] Preferably, in this embodiment, Defect confidence factor Based on the initial aerial screening results, surface risk level, and local anomaly confidence calculation; specifically, based on the aerial imagery and a deep learning detection model, defects are identified in the cells containing candidate points. The highest confidence score in the output is taken and weighted and fused with the normalized risk level value of that cell to obtain the defect confidence factor. , A higher value indicates a higher probability of defects in that area; Structural importance factor The structural importance factor is calculated by combining the region's importance within the overall structure of the site, its protection level, or the attributes of key components. Specifically, based on the cultural relic protection level of the candidate site's location, its load-bearing functional role in the site's structural system (such as load-bearing walls, foundation sections, or auxiliary components), and whether it contains irreplaceable cultural elements (such as inscriptions or murals), each area is quantified, assigned a weighted value, and then weighted and merged to obtain the structural importance factor. , A higher value indicates a higher protection priority for that area.

[0058] Inspection timeliness factor The timeliness factor t is calculated based on the time interval between the last detection and the area, the number of historical missed detections, or the degree of long-term lack of coverage. This includes querying the inspection record database, calculating and normalizing the time interval between the candidate points and the last ground detection, and statistically analyzing the proportion of historical missed detections and the coverage gap in the neighboring area. The three factors are weighted and integrated to obtain the inspection timeliness factor t. The larger the t value, the more outdated the detection information in the area is and the more priority coverage is needed.

[0059] In the fourth step, the ground detection device arrives at the designated locations sequentially according to the detection task sequence. Through pose-synchronized back-projection voxelization and fusion of multi-source surface features, it drives a three-dimensional convolutional neural network to achieve precise detection of cavities inside the site. This embodiment integrates multi-source heterogeneous data—visible light data, point cloud data, and ultrasonic data—into a single three-dimensional voxel model through precise spatial registration, providing high-fidelity, quantifiable digital evidence for subsequent restoration decisions.

[0060] In this step, after the ground detection device reaches the location, the attitude of the robotic arm probe in the ground detection device is first adjusted according to the surface normal vector. The steps include: Obtain the surface normal vector of the target point; that is, perform plane fitting on the local point cloud neighborhood where the target point is located, or interpolate from the triangular mesh to obtain the surface normal vector of the point. ; The direction opposite to the normal vector is taken as the axis vector of the robotic arm probe. ,have By bringing the probe close to the detection surface along the surface normal direction, tangential slippage can be reduced. Based on the axis vector of the robotic arm probe With the desired direction vector Determine the included angle error and the axis of rotation separately, including the following steps; For the robot arm probe axis vector With the desired direction vector Normalize them separately;

[0061] Calculate the included angle error and the axis of rotation using the following formula:

[0062]

[0063] in, Indicates the included angle error. Indicates the axis of rotation.

[0064] Based on the angle error and rotation axis, an attitude increment rotation matrix or quaternion is generated to iteratively correct the pose of the robotic arm end effector until the angle between the probe axis and the surface normal is less than a preset threshold.

[0065] In a preferred embodiment, the robotic arm probe in the ground detection device is moved to a preset distance outside the surface normal, and then slowly fed along the normal. When the contact force sensor detects that the contact force has reached a set threshold, it switches to impedance control mode to maintain a constant contact pressure and complete the detection. This application uses visual servoing and impedance / force control strategies to control the contact state between the probe and the surface of the site, collecting internal detection data while ensuring the safety of the cultural relics, and simultaneously recording the position and attitude information of the probe to provide spatial constraints for subsequent data fusion.

[0066] The contact pressure is:

[0067] In the formula, For contact force, For the desired position, For the desired speed, For actual location, For actual speed, 、 These are control parameters used to maintain a constant contact pressure and reduce mechanical stress on the site.

[0068] In this embodiment, the contact force is controlled within the material's allowable range (preferably <10 N).

[0069] To further optimize the above technical solution, this application uses pose-synchronized back-projection voxelization and surface multi-source feature fusion to drive a three-dimensional convolutional neural network to achieve accurate detection of internal cavities in archaeological sites; the steps include: S1. Collect probe echo data through the robotic arm probe in the ground detection device, including ultrasound / Waveform, GPR time window, thermal image frame, and record probe pose and timestamp, where probe pose includes position and normal vector.

[0070] S2. Based on the pose information and depth information of the echo data, the internal response is projected onto a preset three-dimensional voxel grid, and an initial defect probability voxel map is generated by Bayesian or simple weighted cumulative defect probability. This application adopts conventional Bayesian filtering or occupied grid mapping algorithm to gradually update the log probability of the penetrated voxel according to the distance and angle observation model of the sensor.

[0071] S3. Spatial alignment and feature stitching are performed on the digital surface model and the initial defect probability voxel map in a multi-scale voxel grid, outputting multi-dimensional feature vectors for each voxel, including surface texture, geometric information, material density estimation, and defect probability. Specifically, the surface point cloud is projected and aligned to the aforementioned global voxel grid using a coordinate transformation matrix, and then the RGB and normal vector features of the point cloud are assigned to the corresponding voxels using nearest neighbor interpolation and other methods. Finally, in the feature channel dimension of the voxels, {surface texture, geometric information, material density, and defect probability} are stitched into a multi-dimensional feature vector, which serves as the input to the subsequent 3D convolutional neural network model. The surface texture and geometric information are obtained from the mesh 3D graphics calculated by the UAV, and the material density estimation is obtained through soil sampling, rock hardness testing, etc. It should be noted that the multi-scale voxel mesh in this application originates from point cloud data obtained by surface depth sensor scanning. By calculating the spatial boundaries (extreme values ​​of the X, Y, and Z axes) of the overall surface point cloud, a global bounding box is constructed, and the voxel mesh is divided into three-dimensional spaces according to the set voxel edge lengths to establish the basic framework of the multi-scale voxel mesh.

[0072] S4. The fused 3D digital twin model is divided into voxel blocks and input into a pre-trained 3D convolutional neural network model, outputting a voxel-level hole probability distribution. The hole probability distribution is thresholded, and a clustering algorithm is used to merge voxel clusters and generate a minimum axis-aligned bounding box, which is exported as the model completion task annotation (including position, size, confidence, and suggestion priority). Preferably, the clustering parameters can be set as follows: DBSCAN's eps is 1–3 voxels, minsamples is around 10, and the hole probability threshold is 0.5 by default. In practice, these parameters can be adjusted according to the engineering scenario.

[0073] In some implementations, the 3D convolutional neural network model adopts 3D semantic segmentation / classification structures such as 3D-U-Net, 3D-ResNet or 3D-DenseNet, and uses data augmentation and transfer learning during training to improve generalization ability; The training data is obtained by synthesizing holes in a high-quality model and injecting noise. Optionally, a training set generator is used to synthesize hole samples on the high-quality model and inject sensor noise to form labeled voxel training samples. The training set generator uses a 3D modeling tool (such as the Blender Python API) to randomly place anchor points on the surface of the archaeological site model and generate physical holes through Boolean subtraction. At the same time, Gaussian random offsets are superimposed at the hole boundaries to simulate the measurement errors of photogrammetry and LiDAR, thereby outputting semantically labeled point cloud / voxel training samples.

[0074] In a preferred embodiment of this invention, when the confidence level in the detection result is lower than a preset quality threshold or a valid void is detected, a supplementary testing instruction or a go-around mission plan is automatically generated; and the generated void detection results and mold repair markings are sent to the maintenance work order. Simultaneously, all raw imagery, point cloud data, probe echo data, probe pose data, and voxelized model data are stored with unified geographic coordinates and associated with timestamps to support long-term degradation trend monitoring. The voxelized 3D model supports multi-scale (coarse-to-fine) resolution representation and supports Geo-referenced export format.

[0075] This application takes risk as the main line to drive air-ground collaborative inspection, organically couples wide-area aerial perception with fine ground detection, and uses voxelized space as a unified data carrier to realize the integrated expression and closed-loop optimization of surface information and internal structural information of large linear cultural sites.

[0076] By employing a process of "airborne initial screening - risk assessment - ground-based fixed-point detection - multi-source fusion - automatic cavity detection and feedback", the traditional inspection system can overcome the problems of fragmented surface observation and internal detection, inefficient resource allocation, and high maintenance costs of 3D models, thus forming a scalable, reusable, and sustainably updated digital twin inspection framework for archaeological sites.

[0077] Based on the same inventive concept, embodiments of the present invention also provide a risk-driven, large-scale linear archaeological site air-ground collaborative hierarchical inspection system, comprising: Aerial wide-area perception submodule: includes a drone equipped with a multi-lens oblique photography module, a lightweight LiDAR, and an RTK / IMU positioning unit, used to collect multi-angle images and point clouds along a preset route and generate a digital surface model; in this embodiment, the drone can be a small helicopter, a multi-rotor drone, an unmanned helicopter, or other aerial platforms; The air-ground collaborative task scheduling submodule includes a ground station / edge server, which is used for risk assessment, task scoring and path optimization, and wireless communication with the robot to distribute task queues and paths. Ground contact detection submodule: includes an obstacle-crossing mobile platform equipped with a ≥5-DOF robotic arm, a visual servo camera, force / torque sensors, and flexible probes (ultrasonic or GPR, etc.) for internal detection of ground detection candidate areas; the obstacle-crossing mobile platform is preferably a quadruped, tracked, or biomimetic obstacle-crossing mobile platform, integrating a robotic arm and multi-DOF flexible probes to adapt to the collaborative inspection needs in complex terrain; at the same time, it can optionally integrate one or more of ultrasonic flaw detection, ground penetrating radar (GPR), infrared thermal imaging, or terahertz imaging to achieve multimodal internal structure detection; Digital twin and cavity detection module for archaeological sites: Used for high-performance computing nodes for voxel processing, 3D-CNN inference and DBSCAN clustering, and provides visualization and export interfaces.

[0078] In this embodiment, since the principles by which the above modules solve problems are consistent with the steps in the aforementioned risk-driven, large-scale linear archaeological site air-ground collaborative hierarchical inspection method, the repetitive parts will not be repeated. Please refer to the previous description for details.

[0079] This application utilizes wide-area perception via unmanned aerial vehicles (UAVs) to acquire surface information of archaeological sites and conduct risk assessments, using the risk results to drive the rational allocation of ground inspection resources. It employs ground robots to perform flexible contact-based internal inspections and achieves multi-source data fusion within a voxel-based framework. Simultaneously, it introduces an automatic 3D cavity detection and annotation mechanism, thereby improving inspection efficiency, data integrity, and the ability to assess and maintain archaeological structures. Compared to existing technologies, the core advantages are: 1. A risk-driven, air-ground collaborative, tiered inspection mechanism; For the first time, the results of the risk assessment of the site surface were directly used to drive the generation and scheduling of ground contact detection tasks, realizing the transformation from "full coverage inspection" to "detailed inspection of key areas", and significantly improving the utilization efficiency of limited ground detection resources in large-scale linear sites.

[0080] 2. A data fusion method based on probe pose back projection; By synchronously recording the probe pose information during ground inspection, the internal inspection data is back-projected into the interior of the 3D model according to the physical spatial relationship, and fused with the surface data under the voxelization framework, so as to realize the unified expression of surface geometry, texture information and internal defect information, and solve the problem of the disconnect between surface and internal information in the existing technology.

[0081] 3. Voxelized digital twins serve as a unified carrier of multi-source information; Using voxel grids instead of point clouds or grid models as the basic unit for multi-source data fusion and analysis enables the surface state, internal features, and historical detection results of site structures to be accumulated, compared, and analyzed at the same spatial scale, providing a stable data foundation for long-term monitoring.

[0082] 4. Automatic cavity detection and resurvey closed loop for aerial survey 3D models; To address the issues of voids in large-scale archaeological site aerial survey models and the high cost of manual model repair, a three-dimensional convolutional neural network is introduced to automatically identify voids in voxelized models and generate executable model repair or survey repair task annotations, thus constructing an automated closed loop for model generation and maintenance.

[0083] 5. Cultural relic-friendly flexible contact detection and safety control strategies; In the ground inspection process, a contact strategy combining visual guidance and flexible force control is adopted to effectively reduce mechanical disturbance to the artifacts while ensuring the quality of the inspection data, thus balancing inspection accuracy and the safety of artifact preservation.

[0084] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0085] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites, characterized in that: include: Aerial images and point cloud data of linear archaeological sites to be inspected are collected to generate digital surface models; Based on at least one of terrain, structural, material and environmental indicators, calculate the risk value of each cell in the digital surface model. Candidate areas for ground detection are determined based on the risk values; The ground detection candidate areas are scored for accessibility, and sorted according to the scoring results to generate a detection task sequence; The ground detection device arrives at the designated locations sequentially according to the detection task sequence. Through pose-synchronized back-projection voxelization and fusion of multi-source surface features, it drives a three-dimensional convolutional neural network to achieve accurate detection of cavities inside the site.

2. The risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites as described in claim 1, characterized in that, The topographic indicators are local slope values, local curvature values, or a weighted value of both. The structural index is the crack density in the texture anomaly region; The material index is a weighted value of the normalized degradation index calculated based on the image and the percentage of abnormal area of ​​reflection intensity determined based on the point cloud data. The environmental index is a weighted average of the normalized water index and the normalized vegetation index.

3. The risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites as described in claim 1, characterized in that, The risk value of each cell in the digital surface model is calculated using the following formula: In the formula, Indicates the first The risk value of each unit. This represents the risk calculation function. Indicates topographical indicators, Indicates structural indicators, Indicates material specifications, Indicates environmental indicators.

4. The risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites as described in claim 1, characterized in that, The accessibility scoring of the ground detection candidate areas includes: In the formula, Indicates the risk value. The weights representing risk values Indicates accessibility score, This indicates the weight of the accessibility score. Indicates the detection value score. This indicates the weight of the detection value score.

5. The risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites as described in claim 4, characterized in that, The accessibility score is obtained through the following steps: The path cost is determined as follows; in, Represents path cost, This represents the total path length from the ground detection device to the current point in the ground detection candidate area. This represents the average slope cost along the path. This represents the cost of obstacle density in the path. This indicates the estimated energy consumption along the path. These are the fusion weights for total path length, average slope cost, obstacle density cost, and energy consumption, respectively. A reachability score is determined based on the path cost; In the formula, Accessibility is scored.

6. The risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites as described in claim 4, characterized in that, The calculation process for the detection value score is as follows: In the formula, Indicates the defect confidence factor. The weights corresponding to the defect confidence factors. Indicates structural importance factor, The weights corresponding to the structural importance factors are... Indicates the inspection timeliness factor. The weights corresponding to the inspection timeliness factor.

7. The risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites as described in claim 1, characterized in that, When the ground detection device arrives at the designated points sequentially according to the detection task sequence, the attitude of the robotic arm probe in the ground detection device is adjusted based on the surface normal vector. The steps include: Obtain the normal vector of the surface at the target point; The direction opposite to the normal vector is taken as the axis vector of the robotic arm probe; The included angle error and rotation axis are determined based on the probe axis vector and the desired direction vector of the robotic arm, respectively. An attitude increment rotation matrix is ​​generated based on the angle error and the rotation axis, which is used to iteratively correct the pose of the robotic arm end effector until the angle between the probe axis and the surface normal is less than a preset threshold.

8. The risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites as described in claim 1, characterized in that, When the ground detection device arrives at the designated points sequentially according to the detection task sequence, it performs impedance force control, wherein the contact pressure is: In the formula, For contact force, For the desired position, For the desired speed, For actual location, For actual speed, 、 These are control parameters used to maintain a constant contact pressure and reduce mechanical stress on the site.

9. The risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites as described in claim 1, characterized in that, A 3D convolutional neural network is used to accurately detect internal cavities in archaeological sites by using pose-synchronized back-projection voxelization and fusion of multi-source surface features; including: S1. Collect probe echo data through the robotic arm probe in the ground detection device, and record the probe pose and timestamp; S2. Based on the pose information and the depth information of the echo data, the internal response is projected onto a preset three-dimensional voxel grid, and the probability update algorithm is used to accumulate and calculate the defect probability of penetrating voxels to generate an initial defect probability voxel map. S3. Spatially align and feature-stitch the digital surface model and the initial defect probability voxel map in a multi-scale voxel grid to construct a three-dimensional digital twin model containing surface texture, geometric information, material density estimation and defect probability. S4. Divide the three-dimensional digital twin model into voxel blocks and input them into a pre-trained three-dimensional convolutional neural network model to output a voxel-level hole probability distribution; perform thresholding on the hole probability distribution and use a clustering algorithm to merge connected components to generate hole detection results that include location, size, confidence, and priority.

10. The risk-driven, air-ground collaborative hierarchical inspection method for large-scale linear archaeological sites as described in claim 1 or 9, characterized in that... When the confidence level in the test results is lower than the preset quality threshold or a valid void is detected, a retest instruction or a go-around mission plan is automatically generated; and the generated void detection results and mold repair annotations are sent to the maintenance work order. Simultaneously, all raw images, point cloud data, probe echo data, probe pose data, and voxelized model data are stored with unified geographic coordinates and associated with timestamps to support long-term degradation trend monitoring.

Images