A UAV-based autonomous sensing and inspection method and system for highway slope defects

Abstract

This invention relates to the field of highway disease monitoring technology, and particularly to a method and system for autonomous perception and inspection of highway slope diseases based on unmanned aerial vehicles (UAVs). The method includes: using a UAV to collect time-series image streams and flight attitude data; constructing a three-dimensional spatial topology map based on the time-series image streams and flight attitude data; performing closed-loop control of the UAV's flight attitude according to the three-dimensional spatial topology map; dynamically adjusting the UAV's flight altitude and image acquisition frequency to generate a de-blurred image set and performing illumination invariant enhancement processing on the de-blurred image set; obtaining disease category labels and structural masks from the images in the illumination invariant enhancement-processed de-blurred image set using a topological segmentation method that fuses semantic features; mapping the structural masks onto the three-dimensional spatial topology map, and outputting a disease distribution map with spatial coordinates. This invention effectively solves the technical problem of difficulty in obtaining the spatiotemporal evolution patterns of centimeter-level highway slope diseases through manual inspection.

Description

Technical Field

[0001] This invention relates to the field of highway disease monitoring technology, and in particular to a method and system for autonomous sensing and inspection of highway slope diseases based on unmanned aerial vehicles (UAVs). Background Technology

[0002] As a key geological structure in road engineering, highway slopes are characterized by high concealment and gradual development of defects such as apparent cracking, rock spalling, and local collapse. This continuous development of damage over time and space makes it difficult for manual observation to capture early signs of micro-deformation.

[0003] Current mainstream monitoring methods can be divided into two categories: The first category is based on manual on-site inspections, where technicians use contact measuring tools to measure the size of defects at fixed points. This not only faces safety risks in steep areas, but also cannot cover the entire slope area due to discrete sampling detection. Furthermore, it is limited by insufficient accuracy of visual identification and lacks a systematic understanding of the spatial distribution patterns of crack networks. The second category uses traditional UAV remote sensing. While image acquisition based on pre-set flight paths expands the observation range, it has inherent limitations. Fixed flight altitudes result in image resolution that cannot meet the needs of capturing multi-scale defect features, and motion blur severely affects image usability, forcing equipment to sacrifice flight speed to ensure image quality. More importantly, existing methods exhibit fragmented defects in post-processing of image data. Defect identification only performs isolated annotation on single frames of images, severing the dynamic temporal correlation between images. Survey data and defect semantics cannot be integrated under the constraints of real geological structures, causing existing technologies to remain at the level of recording the instantaneous state of defects. They cannot analyze the correlation between crack propagation paths and spalling areas, nor can they predict the evolution trend of local instability-induced chain disasters. Ultimately, this leads to highway disaster prevention and control decisions lagging behind the actual geological deformation process.

[0004] The information disclosed in this background section is intended only to enhance the understanding of the general background of this disclosure and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0005] This invention provides a method and system for autonomous perception and inspection of highway slope defects based on unmanned aerial vehicles (UAVs), which can effectively solve the problems in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for autonomous perception and inspection of highway slope defects based on unmanned aerial vehicles (UAVs), the method comprising: Use drones to collect time-series image streams of the road area to be inspected and flight attitude data at the corresponding collection time. An updatable three-dimensional spatial topology map is constructed based on the temporal image stream and the flight attitude data; Closed-loop control is performed on the flight attitude of the UAV based on the real-time positioning deviation of the three-dimensional spatial topology map. The flight altitude and image acquisition frequency of the UAV are dynamically adjusted based on a preset ground resolution threshold to generate a set of de-blurred images and perform illumination invariance enhancement processing on the set of de-blurred images. Based on the topological segmentation method that integrates semantic features, the disease category label and structural mask are obtained from the images of the motion-blurred image set after the illumination invariance enhancement process. The structural mask is mapped onto the three-dimensional spatial topology map to output a disease distribution map with spatial coordinates.

[0007] Furthermore, an updatable 3D spatial topology map is constructed, including: Obtain a hierarchical spatial feature point set including terrain contours and surface textures from the time-series image stream, and synchronize the acquisition timestamp of the flight attitude data with the hierarchical spatial feature point set in time and space. The inter-frame spatial pose change is calculated based on the hierarchical spatial feature point set of continuous time frames, and the flight attitude data with corresponding acquisition timestamps are fused to dynamically generate a 3D environmental point cloud with confidence. Based on global relocation optimization, the topology of the 3D environmental point cloud is corrected to generate a sparse semantic map with connectivity. The sparse semantic map is incrementally associated with the arriving temporal image stream to update the geometric boundaries of slope obstacles and terrain undulation information, thereby generating the updatable three-dimensional spatial topology map.

[0008] Furthermore, closed-loop control of the UAV is performed based on a three-dimensional spatial topology map, including: The spatial Euclidean distance deviation between the UAV's projected position in the sparse semantic map and the current planned waypoint in the three-dimensional spatial topology map is calculated in real time. When the spatial Euclidean distance deviation exceeds the pose optimization tolerance, multi-degree-of-freedom attitude compensation parameters are generated, including pitch angle adjustment and yaw angle correction. The attitude compensation parameters are input to the flight controller of the UAV to perform multi-axis cooperative compensation control; The real-time pose information of the UAV after compensation and the newly generated de-motion-blurred image set are collected simultaneously. The effectiveness of pose compensation is verified based on the road reference recognition results in the de-blurred image set, and the current planned waypoint is updated to complete the closed-loop feedback.

[0009] Furthermore, topological segmentation methods that integrate semantic features include: Multi-level convolutional feature extraction is performed on the motion-blurred image set after the illumination invariance enhancement process to generate a semantic feature map set including local texture and global structure. The semantic feature map is input into an attention-guided feature fusion network to adaptively enhance the edge response of the diseased area and generate an initial segmentation mask with topological connectivity. Based on the continuity constraints of the road slope structure, topological relationship optimization is performed on the initial segmentation mask to eliminate isolated noise segmentation regions and repair the fractured geometric boundaries of the defects. The corresponding geological hazard type label is marked on the optimized initial segmentation mask, and the hazard category label and the structural mask that conforms to the continuity of the road slope structure are output simultaneously.

[0010] Further, obtain the disease category markers and structural masks, including: Based on the independent disease regions in the optimized initial segmentation mask, the spatial context feature vector of the independent disease regions in the semantic feature map set is obtained, and a disease feature descriptor including local morphology and global distribution characteristics is generated. The disease feature descriptors are matched with the typical feature distributions of historical disease sample databases to determine the initial classification probability distribution of each independent disease region. Based on the road slope structure continuity constraint, the initial classification probability distribution of adjacent disease areas is spatially consistent to optimize the classification ambiguity and aggregate the fragmented independent disease areas. The optimized initial classification probability distribution is weighted by category decision to generate the disease category label, and the topology of each spatially continuous independent disease region is converted into a vector geographic format to output the structure mask.

[0011] Furthermore, a category decision weighting is performed on the optimized initial classification probability distribution, including: The independent disease areas with spatial adjacency within the roadside slope area are constructed into a disease evolution and propagation subgraph, and the topological feature vector of the disease evolution and propagation subgraph is obtained. Calculate the inter-class association transition probability matrix of the independent disease region based on the temporal evolution pattern of the historical disease sample database, and establish a dynamic coupling relationship with the topological feature vector; Spatially constrained conditional stochastic reasoning is performed on the disease evolution and propagation subgraph to apply context-aware weighting to the probability distribution of the independent disease regions; Based on the stability verification of disease evolution trends, a reliability index for generating category decisions is generated, and the peak regions of probability distribution that satisfy the geological evolution law are selected to output the disease category label.

[0012] Further, mapping the structural mask onto the three-dimensional spatial topology map includes: Multi-level boundary vertex sampling and extraction are performed on the structure mask to generate geometric feature point clouds of each independent disease area in the image coordinate system; Based on the spatial index relationship constructed by the three-dimensional spatial topology map, the geometric feature point cloud and the corresponding spatial location elevation terrain grid are subjected to normal vector constraint registration calculation. The spatial projection drift error of the geometric feature point cloud of the independent disease area between continuous terrain grids is compensated based on the optical flow trajectory of the de-blurred image set and the pose change of the flight attitude data. After spatial attribute binding of the registered and optimized geometric feature point cloud of the independent disease region with the disease category label, topological reconstruction is performed to output the disease distribution map with spatial coordinates that integrates disease geometric morphology and semantic information.

[0013] Further, after spatial attribute binding of the registered and optimized geometric feature point cloud of the independent disease region with the disease category label, topological reconstruction is performed, including: Based on the semantic association rules of the disease category labels, a spatial topological dependency graph is established among the independent disease regions, and a disease attribute relationship network reflecting the evolution and correlation of geological disasters is constructed. The distribution characteristics of the directional gradient histogram of the geometric feature point cloud are obtained and structurally coupled with the disease attribute relationship network to verify the consistency parameters of the three-dimensional deformation trend of each independent disease region. Based on the rock strata structure features of the three-dimensional spatial topology map, the disease attribute relationship network is optimized by performing geomechanical constraints to correct the pseudo-disease point cloud data that does not match the actual terrain. The optimized disease attribute relationship network is implanted into the disease distribution map to generate a disease risk evolution model with spatial hierarchical visualization capabilities.

[0014] A drone-based autonomous sensing and inspection system for highway slope defects, the system comprising: The information acquisition module uses drones to collect time-series image streams of the road area to be inspected and flight attitude data at the corresponding acquisition time. The spatial topology module constructs an updatable 3D spatial topology map based on temporal image streams and flight attitude data; The flight control module performs closed-loop control of the UAV's flight attitude based on the real-time positioning deviation of the three-dimensional spatial topology map. The image processing module dynamically adjusts the UAV's flight altitude and image acquisition frequency based on a preset ground resolution threshold, generates a set of de-blurred images, and performs illumination invariance enhancement processing on the de-blurred image set. The disease labeling module obtains disease category labels and structural masks from the images of the motion-blurred image set after illumination invariance enhancement processing based on the topological segmentation method that integrates semantic features. The map generation module maps the structural mask onto a three-dimensional spatial topological map, outputting a disease distribution map with spatial coordinates.

[0015] Furthermore, the disease marking module includes: The feature convolution unit performs multi-level convolutional feature extraction on the motion-blurred image set after illumination invariance enhancement, generating a semantic feature map set including local texture and global structure. The initial segmentation unit inputs the semantic feature map into the attention-guided feature fusion network, which adaptively enhances the edge response of the diseased area and generates an initial segmentation mask with topological connectivity. Geometric topology units, based on the continuity constraints of road slope structures, perform topological relationship optimization on the initial segmentation mask to eliminate isolated noise segmentation regions and repair fractured geometric boundaries; The label matching unit marks the corresponding geological hazard type label on the optimized initial segmentation mask, and simultaneously outputs the hazard category label and the structural mask that conforms to the continuity of the road slope structure.

[0016] The technical solution of this invention can achieve the following technical effects: By integrating 3D spatial modeling and semantic topology segmentation technologies with an autonomous closed-loop control system for unmanned aerial vehicles, a spatial evolution model of road slope defects with geomechanical constraints is constructed in real time, solving the technical problem that manual inspections cannot obtain the spatiotemporal evolution patterns of centimeter-level road slope defects.

[0017] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0018] 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 some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a method for autonomous perception and inspection of highway slope defects based on unmanned aerial vehicles (UAVs). Figure 2A flowchart illustrating the process of constructing a 3D spatial topology map; Figure 3 A flowchart illustrating a topological segmentation method that integrates semantic features; Figure 4 This is a schematic diagram illustrating the process of mapping a structural mask to a three-dimensional spatial topology map. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] Example 1; like Figure 1 As shown, this application provides an autonomous sensing and inspection method for highway slope defects based on unmanned aerial vehicles (UAVs). The method includes: Use drones to collect time-series image streams of the road area to be inspected and flight attitude data at the corresponding collection time. An updatable 3D spatial topology map is constructed based on temporal image streams and flight attitude data; Closed-loop control of the UAV's flight attitude is performed based on the real-time positioning deviation of the three-dimensional spatial topology map. The flight altitude and image acquisition frequency of the UAV are dynamically adjusted based on a preset ground resolution threshold to generate a set of de-blurred images and perform illumination invariance enhancement processing on the set of de-blurred images. Based on the topological segmentation method that integrates semantic features, disease category labels and structural masks are obtained from the images of the motion-blurred image set after illumination invariance enhancement. The structural mask is mapped onto a three-dimensional spatial topology map, and the disease distribution map with spatial coordinates is output.

[0023] Specifically, the process begins with deploying a rotary-wing UAV equipped with a multi-axis stabilization gimbal and RTK positioning module. Upon startup, an initial 3D terrain model of the target road area is loaded as a baseline. During the autonomous inspection phase, the UAV flies along the slope at a preset initial altitude and speed, acquiring a temporal image stream using the onboard camera's global exposure mode. For example, it acquires 5 frames of 20-megapixel images per second, simultaneously recording the POS data for each frame, including latitude, longitude, elevation, pitch angle, and roll angle. To overcome the limitations of traditional static modeling, a spatiotemporal synchronization field is constructed for the acquired image stream and attitude data. The timestamps of the image frames and POS data are aligned at the millisecond level. Simultaneously, combined with camera intrinsic parameters, the pixel coordinate system of each image frame is transformed to a world coordinate system based on the slope's rock strata orientation. Incremental 3D reconstruction is then performed, selecting three consecutive images as a computational unit and extracting adjacent images using an adaptive feature matching algorithm. The system uses corresponding point pairs, such as matching SIFT feature points with the gradient consistency of rock fracture texture, and then uses motion reconstruction structure technology to iteratively calculate the spatial point positions, adding a newly reconstructed slope point cloud on the basis of the initial 3D model. In particular, for overlapping point clouds in the slope turning area, a normal vector constraint registration strategy is adopted, that is, by calculating the principal curvature direction of the local point cloud, feature plane alignment is performed to eliminate splicing misalignment, generating a high-precision 3D spatial topology map that can be expanded in real time. During the continuous flight of the UAV, the real-time measured image positioning coordinates are compared with the theoretical coordinates of the corresponding positions in the map to generate positioning deviation. For example, if the actual coordinates of a rock block corner in the image are offset by 0.2 meters towards the foot of the slope compared with the map coordinates, a control strategy is triggered based on the magnitude of the deviation vector. When the horizontal offset exceeds the first threshold, such as 0.15 meters, the UAV yaw angle is adjusted to correct the trajectory. When the elevation offset exceeds the second threshold, such as 0.At 1 meter, the pitch angle is compensated by a gimbal stabilization mechanism to ensure the vertical viewing angle of the image. Ground resolution thresholds are set according to the actual needs of disease monitoring; for example, crack monitoring requires a resolution of 5 mm / pixel, and spalling area monitoring requires 10 mm / pixel. During flight, the actual resolution of the current position is calculated in real time. Based on the geometric relationship between the camera focal length and the current flight altitude, and through pre-recognition of preceding images, when a dense area of ​​rock fissures is detected, the flight altitude is automatically reduced to the threshold height, for example, from 30 meters to 12 meters. Simultaneously, the acquisition frequency is increased proportionally, for example, from 5 frames / second to 8 frames / second. To counteract velocity fluctuations caused by altitude changes, optical flow analysis is performed on adjacent image frames to extract pixel motion vector fields and infer camera jitter trajectories. An inverse filter is then used to reconstruct and eliminate the influence of the blur kernel, generating a de-blurred image set that includes slope details. To address fluctuations in field illumination, adaptive histogram equalization is applied to the deblurred images, maintaining illumination consistency between image groups while preserving rock strata information in cracked dark and bright areas. A pre-trained semantic segmentation network extracts disease category features from single-frame images, including typical labels such as cracks, spalling, and seepage. Simultaneously, a parallel network branch outputs pixel-level structural masks of the diseased areas. The disease mask of each image frame is projected onto the corresponding triangular mesh patch of the 3D spatial topology map. Based on the curvature change characteristics of the disease boundary after projection, such as the dendritic branching of crack ends in 3D space, the segmentation results in the 2D image are optimized and corrected. Cracks are marked as spatial polyline trajectories with each vertex containing XYZ coordinates, spalling areas are marked as triangular mesh clusters with the thickness of rock spalling, and seepage areas are marked as polygonal surfaces with the gradient values ​​of area and infiltration depth. For example, when a UAV performs inspection on a 70° steep slope, it initially flies at a height of 25 meters to collect images of the cliff top area. When the 3D map reconstruction shows... The image revealed a suspected crack area on the slope, with localized linear depressions in the point cloud. The system immediately descended to a height of 8 meters for detailed image capture. During close-range acquisition, airflow disturbance caused motion blur in frame 205. The pixel displacement path of this frame was reconstructed from adjacent frames 204 and 206, resulting in a clear image. In the identification phase, the crack appeared as an isolated line segment in the 2D image, but after projection onto a 3D map, its extension direction coincided with the rock joint surface. Therefore, it was corrected to a long-range continuous crack, and its length was marked. The final output map visually displays the 3D trajectory of the crack's development along the weak structural surface within the slope.

[0024] The technical solution of this invention integrates three-dimensional spatial modeling and semantic topology segmentation technology based on the UAV autonomous closed-loop control system to construct a disease spatial evolution model with geomechanical constraints in real time, which solves the technical problem that manual inspection is difficult to obtain the spatiotemporal evolution law of centimeter-level diseases on highway slopes.

[0025] Furthermore, such as Figure 2 As shown, constructing an updatable 3D spatial topology map includes: The hierarchical spatial feature point set, including terrain contours and surface textures, is obtained from the temporal image stream, and the acquisition timestamp of the flight attitude data is synchronized with the hierarchical spatial feature point set in time and space. The spatial pose change between frames is calculated based on the hierarchical spatial feature point set of continuous time frames, and the flight attitude data with corresponding acquisition timestamps are fused to dynamically generate a 3D environmental point cloud with confidence. Based on global relocation optimization, topological correction is performed on the 3D environment point cloud to generate a sparse semantic map with connectivity. Incremental feature association is performed between sparse semantic maps and arriving temporal image streams to update the geometric boundaries of slope obstacles and terrain undulation information, generating an updatable 3D spatial topology map.

[0026] As a preferred embodiment of the above, firstly, a drone system is deployed, equipped with a high-performance camera to ensure high resolution and quality of image data. During the inspection process, the drone not only needs to capture the overall terrain layout and surface features of the roadside slope, but also needs to extract a set of hierarchical spatial feature points. This set of feature points includes significant changes in the terrain contour, such as slope boundaries and edge contours, ensuring the drone's comprehensive observation of terrain changes. Each frame of image acquired by the drone is recorded with a timestamp, and these timestamps are precisely aligned with the flight attitude data, enabling the drone to achieve precise positioning of each image data in three-dimensional space. The attitude data includes key pose information such as the drone's pitch and roll angles. The application of spatiotemporal synchronization technology ensures a high degree of consistency between the camera-captured images and the actual spatial layout. Next, by analyzing the hierarchical spatial feature points of consecutive time-series frames, the spatial pose changes between frames are calculated. These changes are fused with the flight attitude data corresponding to the acquisition timestamp to dynamically generate a 3D environmental point cloud with confidence. Combined with the timestamp, the generated point cloud not only displays the physical structure of the slope but also reflects the environmental characteristics at the time of data acquisition. Through confidence assessment, low-quality data can be identified and eliminated, retaining the most reliable point cloud information, thereby significantly improving... The accuracy of the overall spatial map is improved. After generating the initial point cloud, global relocalization optimization is performed to correct the topological structure of the 3D environmental point cloud. This optimization process combines spatial connectivity and terrain semantic features to generate a sparse semantic map with connectivity relationships. In addition to showing the spatial relationships of the point cloud, the sparse semantic map also includes relevant geological semantic information such as rock strata structure and vegetation cover. Such a map helps to comprehensively identify terrain features and potential structural defects, breaking through the limitations of traditional methods that rely solely on visual information. Finally, the sparse semantic map is incrementally correlated with the real-time received temporal image stream. This process enables the flight system to continuously expand its... The dataset dynamically updates the geometric boundaries and undulations of the slope terrain. The application of incremental association methods means that the map is no longer a static image, but a dynamic model that responds to environmental changes in real time. In complex terrain environments, drones can adjust map information and flight strategies in a timely manner as the terrain and obstacles change. For example, during highway slope inspections, drones may encounter sudden terrain changes, such as landslides or new rock fissures. Through continuous image streams and constantly updated sparse semantic maps, drones can quickly identify these changes and make adjustments. This not only improves the accuracy and timeliness of road inspections, but also ensures the safety of the drone's flight path.

[0027] Furthermore, closed-loop control of the drone based on a three-dimensional spatial topology map includes: Real-time calculation of the spatial Euclidean distance deviation between the UAV's projected position in the sparse semantic map and the current planned waypoint in the 3D spatial topology map; When the spatial Euclidean distance deviation exceeds the pose optimization tolerance, multi-degree-of-freedom attitude compensation parameters are generated. These parameters include pitch angle adjustment and yaw angle correction. The attitude compensation parameters are input into the UAV's flight controller to perform multi-axis cooperative compensation control; Simultaneously acquire the real-time pose information of the compensated UAV and the latest generated de-motion-blurred image set; The effectiveness of pose compensation is verified based on the road reference recognition results in the de-blurred image set, and the current planned waypoints are updated to complete the closed-loop feedback.

[0028] As a preferred embodiment of the above, during operation, the UAV calculates in real time the spatial Euclidean distance deviation between its projected position in the sparse semantic map and the real-time planned waypoint. If this deviation exceeds a preset pose optimization tolerance, multi-degree-of-freedom attitude compensation parameters need to be generated. These attitude compensation parameters include key pitch angle adjustment and heading angle correction to ensure that the UAV can quickly recover to the optimal flight path. Once the compensation parameters are generated, they are input to the UAV's flight controller for multi-axis collaborative compensation control. The UAV corrects its flight path by finely adjusting its own attitude. In implementation, flight control and attitude adjustment are combined to ensure flight stability and accuracy. This step achieves automation and precision in flight attitude control by ensuring the UAV's real-time correction capability when the flight path deviates. After attitude compensation, the UAV continues to collect the compensated real-time pose information and the latest generated de-motion-blurred image set. For these images, the UAV further identifies road references and verifies the effectiveness of pose compensation. By comparing the positions of the references in the images with the theoretical positions, it is determined whether the UAV has recovered through compensation. The system returns to the correct flight path; if the compensation result is verified to be effective, the current planned flight path is updated to ensure closed-loop feedback. This update mechanism ensures that every compensation decision is accurately recorded and applied, forming a continuously iterative closed-loop control loop. The main innovation of this real-time adjustment and closed-loop feedback mechanism is that it significantly improves the reliability and accuracy of UAV inspections in complex terrain. In practical applications, this implementation scheme can effectively cope with UAV flight deviations caused by factors such as wind speed changes and irregular terrain obstacles. In highway slope inspections, if the UAV encounters sudden climate changes or irregular ground obstacles, this closed-loop control scheme can proactively adjust the flight path and implement compensation to continue to maintain accurate inspection of slope defects. For example, in complex mountainous environments, this technology can ensure that the UAV flies along the predetermined trajectory without being significantly affected by sudden terrain changes or airflow disturbances, ensuring the integrity and reliability of inspection data. This intelligent and precise UAV control brings a revolutionary improvement to highway slope inspection compared to traditional methods, helping to achieve more efficient road safety management.

[0029] Furthermore, such as Figure 3 As shown, the topology segmentation method that integrates semantic features includes: Multi-level convolutional feature extraction is performed on the motion-blurred image set after illumination invariance enhancement to generate a semantic feature map set including local texture and global structure. The semantic feature map is input into the attention-guided feature fusion network to adaptively enhance the edge response of the diseased area and generate an initial segmentation mask with topological connectivity. Based on the continuity constraints of the road slope structure, topological relationship optimization is performed on the initial segmentation mask to eliminate isolated noise segmentation regions and repair the fractured geometric boundaries of the defects. Label the corresponding geological hazard type on the optimized initial segmentation mask, and simultaneously output the hazard category marker and the structural mask that conforms to the continuity of the road slope structure.

[0030] As a preferred embodiment of the above, firstly, multi-level convolutional feature extraction is performed on the motion-blurred image set after illumination invariance enhancement to generate a semantic feature map including local texture and global structure. Multi-level convolutional processing significantly improves the accuracy of disease area identification by enhancing local details and global morphology in the image. Local texture extraction can capture surface features such as fine cracks and peeling, while global structure analysis helps to understand the relationship between the disease and the surrounding environment. This hierarchical feature extraction technique ensures that no important diseases are missed by comprehensively analyzing information from different dimensions. Subsequently, the semantic feature map is input into an attention-guided feature fusion network. This network adaptively enhances the edge response of the disease area and generates an initial segmentation mask with topological connectivity. The attention mechanism utilizes the edge information of the disease area to strengthen the identification ability of the disease boundary, thereby making the boundary identification clearer and more accurate. The initial segmentation mask ensures the spatial continuity of the identified disease area by emphasizing the topological connectivity of the disease area. To ensure integrity, an advanced attention-guided algorithm further enhances segmentation accuracy, enabling disease features to be prominently displayed in the image. Next, based on the continuity constraints of the road slope structure, the initial segmentation mask undergoes topological optimization. This optimization automatically eliminates isolated noisy segmentation regions and repairs fractured disease geometric boundaries without manual intervention. This strategy ensures that the geometric representation of the disease conforms to the actual structural characteristics of the road slope. Continuity constraints allow the disease to more realistically reflect its natural distribution in the image, reducing segmentation errors caused by environmental noise. Finally, the optimized initial segmentation mask is labeled with corresponding geological disease type tags, and simultaneously outputs disease category labels and a structural mask conforming to the continuity of the road slope structure. The geological disease type tags are automatically matched and labeled based on a rich disease database, achieving standardization and efficiency in disease classification. The final generated structural mask can be used for further analysis and disease location, providing reliable data for subsequent disaster prevention and on-site maintenance.

[0031] Furthermore, obtaining disease category markers and structural masks includes: Based on the independent disease regions in the optimized initial segmentation mask, the spatial context feature vectors of the independent disease regions in the semantic feature map are obtained, and disease feature descriptors including local morphology and global distribution characteristics are generated. The initial classification probability distribution of each independent disease region is determined by performing similarity matching calculations between the disease feature descriptors and the typical feature distributions of the historical disease sample database. Based on the structural continuity constraints of the road slope, the initial classification probability distribution of adjacent disease areas is spatially consistent to optimize the classification ambiguity and aggregate fragmented independent disease areas. The optimized initial classification probability distribution is weighted by category decision to generate disease category labels, and the topology of each spatially continuous independent disease region is converted into a vector geographic format output structure mask.

[0032] As a preferred embodiment of the above, firstly, independent disease regions are identified in the optimized initial segmentation mask, and their spatial context feature vectors in the semantic feature map are obtained based on these regions. Specifically, the context feature vector of each independent disease region includes the specific location, morphological features, and distribution characteristics relative to other disease regions in the entire map. Using this information, a disease feature descriptor reflecting the local morphological and global distribution characteristics of the disease region is generated. This process can finely distinguish adjacent or overlapping disease features. Subsequently, by performing similarity matching calculations between the generated disease feature descriptors and the typical feature distributions of historical disease sample databases, the initial classification probability distribution of each independent disease region is determined. The historical data in the sample database provides rich disease classification priors, ensuring the accuracy and efficiency of similarity matching, while the initial classification probability distribution provides a basis for subsequent decisions. Next, in order to improve the accuracy of disease identification, spatial one- Consistency optimization, under the constraint of road slope structural continuity, optimizes the initial classification probability distribution of adjacent disease areas. This optimization process eliminates classification ambiguities between areas, aggregates and unifies fragmented independent disease areas, and ensures the consistency and accuracy of classification results. Through this method, areas that may have been misclassified due to noise or irregular shapes can be corrected and accurately labeled. Finally, category decision weighting is performed on the optimized initial classification probability distribution to generate the final disease category label. According to the importance and distribution of different disease types, category decision weighting can ensure the rationality and accuracy of disease type labeling. Furthermore, the topological structure of spatially continuous independent disease areas is transformed into a vector geographic format and output as a structural mask. This method of transforming into a vector geographic format ensures that the output data can be used for further analysis and geographic information system (GIS) integration, providing direct and effective data support for road inspection decision-making.

[0033] Furthermore, the optimized initial classification probability distribution is weighted according to category decisions, including: Independent disease areas with spatial adjacency within the roadside slope area are constructed as disease evolution and propagation subgraphs, and the topological feature vectors of the disease evolution and propagation subgraphs are obtained. Calculate the inter-class association transition probability matrix of independent disease regions based on the temporal evolution pattern of the historical disease sample database, and establish a dynamic coupling relationship with the topological feature vector; Conditional stochastic reasoning with spatial constraints is applied to the probability distribution of independent disease regions on the disease evolution and propagation subgraph, and context-aware weighting is applied to the probability distribution of independent disease regions. Based on the stability verification of disease evolution trends, a reliability index for generating category decisions is generated, and disease category labels are output by screening the probability distribution peak areas that satisfy the geological evolution laws.

[0034] As a preferred embodiment of the above, firstly, independent disease areas with spatial adjacency within the road slope area are constructed as disease evolution and propagation subgraphs, thereby obtaining the topological feature vectors of the disease evolution and propagation subgraphs. These feature vectors include the spatial arrangement and propagation paths of various slope diseases, revealing the possible evolutionary trends and mutual influences of diseases over time. Through the visualized subgraph structure, a deeper understanding of the interaction relationships between diseases can be achieved, thus providing a solid foundation for subsequent analysis steps. Next, using the temporal evolution patterns stored in the historical disease sample database, the inter-class association transition probability matrix of independent disease areas is calculated. This probability matrix describes the transition probability between different disease types during the evolution process. Furthermore, a dynamic coupling relationship is established with the topological feature vectors. By connecting the evolution patterns and spatial features, real-time assessment of the dynamic distribution of diseases can be achieved. This step provides foresight for predicting disease development. The system is supported by robust data. Spatially constrained conditional stochastic inference is performed on the disease evolution and propagation submap to apply context-aware weighting to the probability distribution of independent disease areas. This method, combined with conditional random field theory, dynamically adjusts the classification probability of each disease area while considering spatial constraints. The context-aware weighting strategy ensures more reliable and precise disease area classification results, reducing misclassifications caused by spatial limitations through the inference process. To guarantee high reliability of category decisions, a reliability index for category decisions is generated based on the stability verification of disease evolution trends. During this process, peak regions of probability distributions that satisfy geological evolution laws are selected to output accurate disease category labels. Through a robust verification process, arbitrariness and inaccurate labels caused by data anomalies can be eliminated to the greatest extent, ensuring that the generated category labels truly reflect the actual situation of slope diseases.

[0035] Furthermore, such as Figure 4 As shown, mapping the structural mask to a three-dimensional spatial topology map includes: Multi-level boundary vertex sampling and extraction is performed on the structural mask to generate geometric feature point clouds of each independent disease area in the image coordinate system; Based on the spatial index relationship constructed by the 3D spatial topology map, normal vector constraint registration calculation is performed between the geometric feature point cloud and the corresponding spatial location elevation terrain grid. Compensate for spatial projection drift error of geometric feature point cloud in independent disease areas between continuous terrain grids based on pose changes of optical flow trajectory and flight attitude data of de-blurred image set; After spatial attribute binding of the registered and optimized geometric feature point cloud of the independent disease area with the disease category label, topological reconstruction is performed to output a disease distribution map with spatial coordinates that integrates disease geometric morphology and semantic information.

[0036] As a preferred embodiment of the above, firstly, multi-level boundary vertex sampling is performed on the generated structural mask. After systematic processing, geometric feature point clouds of each independent disease area in the image coordinate system are generated. In this process, by extracting irregularly shaped vertices on the disease boundary, the boundary detail information of the disease can be captured, ensuring the integrity of the disease geometric description. The sampling process dynamically adjusts the sampling resolution according to the complexity of the disease area. For example, higher sampling frequency may be used for crack edges, while a moderate sampling interval is maintained for peeling areas to balance accuracy and computational efficiency. Next, a spatial index relationship is constructed based on a three-dimensional spatial topological map, and the extracted geometric feature point clouds are compared with the corresponding spatial location elevation topography network. The normal vector constraint registration calculation is performed. This technique analyzes the normal relationship between feature points and the terrain grid surface, enabling the geometric feature point cloud to be mapped more realistically to the actual slope terrain. This registration process corrects misalignment caused by terrain changes, ensuring that the spatial representation of the fault features is highly consistent with the actual geographical location. To further optimize the mapping results, the spatial projection drift error of the geometric feature point cloud between continuous terrain grids is compensated based on the optical flow trajectory of the de-blurred image set and the pose changes of the UAV's flight attitude. The compensation process dynamically corrects drift errors caused by flight jitter, image blur, etc., by analyzing the optical flow trajectory between image frames. For example, when the UAV is in a strong... When airflow causes image shift, combining pose data with optical flow trajectories ensures that the spatial positions of feature points remain consistent within a continuous terrain grid. Then, the registered and optimized geometric feature point cloud is spatially bound to the corresponding disease category markers. This spatial attribute binding process integrates the semantic information of disease types with the geometric details of disease areas into a unified data structure, providing direct support for comprehensive disease analysis. Combined with spatial attribute information, the location, type, and physical characteristics of diseases in the 3D environment can be clearly identified, forming a logically clear and data-rich disease description. Finally, based on the binding results, topological reconstruction is performed on the topology of each disease area, outputting a fusion of disease geometry and semantic information. The method generates a disease distribution map with spatial coordinates. The topology reconstruction process ensures the integrity and consistency of the generated data, while the final output disease distribution map is represented in vector geographic format to show the actual distribution relationship between diseases and terrain in three-dimensional space. This distribution map has both visualization effect and operability, and can be directly used for the integration of subsequent disease assessment and prevention systems. For example, in highway slope inspection tasks, when a drone detects structural diseases, such as rock spalling or through cracks, the disease distribution map output by this method can intuitively show the location, size and development trend of the disease on the slope. Combined with this map, the maintenance team can quickly assess the degree of disease risk, plan targeted intervention measures, and reduce the impact of the disease.

[0037] Furthermore, after spatial attribute binding of the registered and optimized geometric feature point cloud of the independent disease area with the disease category label, topological reconstruction is performed, including: Based on the semantic association rules of disease category labels, a spatial topological dependency graph is established between each independent disease region, and a disease attribute relationship network reflecting the evolution and correlation of geological disasters is constructed. The distribution characteristics of the directional gradient histogram of geometric feature point cloud are obtained and structurally coupled with the disease attribute relationship network for verification, and the consistency parameters of the three-dimensional deformation trend of each independent disease area are determined. Based on the rock strata structure features of the three-dimensional spatial topology map, the geological and mechanical constraints of the disease attribute relationship network are optimized to correct the pseudo-disease point cloud data that does not match the actual terrain. The optimized disease attribute relationship network is embedded into the disease distribution map to generate a disease risk evolution model with spatial hierarchical visualization capabilities.

[0038] As a preferred embodiment of the above, firstly, a spatial topological dependency graph between independent disease areas is established based on the semantic association rules of disease category labels. This spatial topological dependency graph uses disease categories and characteristics as nodes and potential spatial associations as edges. By analyzing the interaction relationships between regions, a disease attribute relationship network reflecting the evolutionary association of geological hazards is constructed. The semantic association rules are a hierarchical framework based on semantic information including disease category labels, topographic characteristics, and disease scale. For example, the expansion paths between fracture areas may exhibit strong node relationships, while the evolutionary dependency relationship between spalling areas and seepage areas may be reflected through erosion or flow channels. The disease attribute relationship network... This method not only integrates spatial data on disease distribution but also provides a foundation for predicting the dynamic trends of geological disasters. Next, based on the distribution characteristics of the directional gradient histogram of geometric feature point clouds, it is structurally coupled and verified with the disease attribute relationship network. The directional gradient histogram characterizes the edge direction distribution characteristics of diseased areas, capturing the detailed morphology and development direction of geological diseases. Through coupling verification, the consistency parameters of the three-dimensional deformation trend of independent diseased areas can be calculated and determined, thereby analyzing the dynamic extension, contraction, and deformation patterns between diseased areas. For example, if the gradient distribution of certain fractured diseased areas indicates a consistent deformation direction, it may be possible to predict their expansion path tendency and adjacency. The regions exhibit evolutionary correlations, and this verification ensures the logical consistency between the three-dimensional morphology and spatial distribution of the diseased areas. Then, under the constraints of the rock strata structure characteristics in the three-dimensional spatial topological map, the disease attribute relationship network is optimized using geomechanical constraints. During this optimization process, the structural characteristics of the rock strata, such as the main lines, the actual direction of the strata, and the distribution of weak surfaces, are utilized to perform mechanical coordination calculations on the disease geometric feature point cloud data. This corrects for pseudo-disease point cloud data that does not conform to the actual terrain conditions. For example, by identifying stress concentration areas, virtual crack markers caused by image errors can be eliminated, thereby significantly improving the accuracy of the geometric point cloud and the authenticity of the data attributes. This mechanical optimization... The process combines spatial rationality with physical constraint logic, providing a reliable basis for refining the distribution of diseases. Finally, the optimized disease attribute relationship network is implanted into the disease distribution map to generate a disease risk evolution model with spatial hierarchical visualization capabilities. During the implantation stage, the information of each node in the disease relationship network, including the three-dimensional spatial location, category label, and evolutionary relationship of the disease, is fully loaded into the disease distribution map. A risk expression framework is constructed based on the dynamic geological change trend. The resulting disease risk evolution model not only has the ability to display static geological distribution but also has dynamic prediction capabilities, revealing the potential threat range and development path of disease expansion through risk level identification.For example, in actual drone inspection missions, when a drone flies along a highway slope, the aforementioned model can be used to capture real-time changes in geological defects, determine risk levels, and identify potential disaster zones. For marked cracks on the slope, if a continuous, expanding gradient change is observed around the crack and the evolution network shows increased stress concentration between related nodes in the area, the model can predict that the crack may further penetrate and cause damage, and issue warnings to the relevant areas.

[0039] Example 2; Based on the same inventive concept as the UAV-based autonomous sensing and inspection method for highway slope defects in the foregoing embodiments, this invention also provides a UAV-based autonomous sensing and inspection system for highway slope defects, the system comprising: The information acquisition module uses drones to collect time-series image streams of the road area to be inspected and flight attitude data at the corresponding acquisition time. The spatial topology module constructs an updatable 3D spatial topology map based on temporal image streams and flight attitude data; The flight control module performs closed-loop control of the UAV's flight attitude based on the real-time positioning deviation of the three-dimensional spatial topology map. The image processing module dynamically adjusts the UAV's flight altitude and image acquisition frequency based on a preset ground resolution threshold, generates a set of de-blurred images, and performs illumination invariance enhancement processing on the de-blurred image set. The disease labeling module obtains disease category labels and structural masks from the images of the motion-blurred image set after illumination invariance enhancement processing based on the topological segmentation method that integrates semantic features. The map generation module maps the structural mask onto a three-dimensional spatial topological map, outputting a disease distribution map with spatial coordinates.

[0040] The adjustment system described above in this invention can effectively realize the autonomous perception and inspection method for highway slope diseases based on UAVs. The technical effects it can achieve are as described in the above embodiments, and will not be repeated here.

[0041] Furthermore, the disease marking module includes: The feature convolution unit performs multi-level convolutional feature extraction on the motion-blurred image set after illumination invariance enhancement, generating a semantic feature map set including local texture and global structure. The initial segmentation unit inputs the semantic feature map into the attention-guided feature fusion network, which adaptively enhances the edge response of the diseased area and generates an initial segmentation mask with topological connectivity. Geometric topology units, based on the continuity constraints of road slope structures, perform topological relationship optimization on the initial segmentation mask to eliminate isolated noise segmentation regions and repair fractured geometric boundaries; The label matching unit marks the corresponding geological hazard type label on the optimized initial segmentation mask, and simultaneously outputs the hazard category label and the structural mask that conforms to the continuity of the road slope structure.

[0042] Similarly, the above-mentioned optimization schemes for the system can also achieve the optimization effects corresponding to the methods in Embodiment 1, which will not be repeated here.

[0043] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for autonomous perception and inspection of highway slope defects based on unmanned aerial vehicles (UAVs), characterized in that, The method includes: Use drones to collect time-series image streams of the road area to be inspected and flight attitude data at the corresponding collection time. An updatable three-dimensional spatial topology map is constructed based on the temporal image stream and the flight attitude data; Closed-loop control is performed on the flight attitude of the UAV based on the real-time positioning deviation of the three-dimensional spatial topology map. The flight altitude and image acquisition frequency of the UAV are dynamically adjusted based on a preset ground resolution threshold to generate a set of de-blurred images and perform illumination invariance enhancement processing on the set of de-blurred images. Based on the topological segmentation method that integrates semantic features, the disease category label and structural mask are obtained from the images of the motion-blurred image set after the illumination invariance enhancement process. The structural mask is mapped onto the three-dimensional spatial topology map to output a disease distribution map with spatial coordinates; Topological segmentation methods that integrate semantic features include: Multi-level convolutional feature extraction is performed on the motion-blurred image set after the illumination invariance enhancement process to generate a semantic feature map set including local texture and global structure. The semantic feature map is input into an attention-guided feature fusion network to adaptively enhance the edge response of the diseased area and generate an initial segmentation mask with topological connectivity. Based on the continuity constraints of the road slope structure, topological relationship optimization is performed on the initial segmentation mask to eliminate isolated noise segmentation regions and repair the fractured geometric boundaries of the defects. The corresponding geological hazard type label is marked on the optimized initial segmentation mask, and the hazard category label and the structural mask that conforms to the continuity of the road slope structure are output simultaneously. Obtain the disease category marker and structure mask, including: Based on the independent disease regions in the optimized initial segmentation mask, the spatial context feature vector of the independent disease regions in the semantic feature map set is obtained, and a disease feature descriptor including local morphology and global distribution characteristics is generated. The disease feature descriptors are matched with the typical feature distributions of historical disease sample databases to determine the initial classification probability distribution of each independent disease region. Based on the road slope structure continuity constraint, the initial classification probability distribution of adjacent disease areas is spatially consistent to optimize the classification ambiguity and aggregate the fragmented independent disease areas. The optimized initial classification probability distribution is weighted by category decision to generate the disease category label, and the topology of each spatially continuous independent disease region is converted into a vector geographic format to output the structure mask; Perform category decision weighting on the optimized initial classification probability distribution, including: The independent disease areas with spatial adjacency within the roadside slope area are constructed into a disease evolution and propagation subgraph, and the topological feature vector of the disease evolution and propagation subgraph is obtained. Calculate the inter-class association transition probability matrix of the independent disease region based on the temporal evolution pattern of the historical disease sample database, and establish a dynamic coupling relationship with the topological feature vector; Spatially constrained conditional stochastic reasoning is performed on the disease evolution and propagation subgraph to apply context-aware weighting to the probability distribution of the independent disease regions; Based on the stability verification of disease evolution trends, a reliability index for generating category decisions is generated, and the peak regions of probability distribution that satisfy the geological evolution law are selected to output the disease category label.

2. The method for autonomous perception and inspection of highway slope defects based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, Construct an updatable 3D spatial topology map, including: Obtain a hierarchical spatial feature point set including terrain contours and surface textures from the time-series image stream, and synchronize the acquisition timestamp of the flight attitude data with the hierarchical spatial feature point set in time and space. The inter-frame spatial pose change is calculated based on the hierarchical spatial feature point set of continuous time frames, and the flight attitude data with corresponding acquisition timestamps are fused to dynamically generate a 3D environmental point cloud with confidence. Based on global relocation optimization, the topology of the 3D environmental point cloud is corrected to generate a sparse semantic map with connectivity. The sparse semantic map is incrementally associated with the arriving temporal image stream to update the geometric boundaries of slope obstacles and terrain undulation information, thereby generating the updatable three-dimensional spatial topology map.

3. The method for autonomous perception and inspection of highway slope defects based on unmanned aerial vehicles (UAVs) according to claim 2, characterized in that, Perform closed-loop control of the UAV based on a 3D spatial topology map, including: The spatial Euclidean distance deviation between the UAV's projected position in the sparse semantic map and the current planned waypoint in the three-dimensional spatial topology map is calculated in real time. When the spatial Euclidean distance deviation exceeds the pose optimization tolerance, multi-degree-of-freedom attitude compensation parameters are generated, including pitch angle adjustment and yaw angle correction. The attitude compensation parameters are input to the flight controller of the UAV to perform multi-axis cooperative compensation control; The real-time pose information of the UAV after compensation and the newly generated de-motion-blurred image set are collected simultaneously. The effectiveness of pose compensation is verified based on the road reference recognition results in the de-blurred image set, and the current planned waypoint is updated to complete the closed-loop feedback.

4. The method for autonomous perception and inspection of highway slope defects based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, Mapping the structural mask onto the three-dimensional spatial topology map includes: Multi-level boundary vertex sampling and extraction are performed on the structure mask to generate geometric feature point clouds of each independent disease area in the image coordinate system; Based on the spatial index relationship constructed by the three-dimensional spatial topology map, the geometric feature point cloud and the corresponding spatial location elevation terrain grid are subjected to normal vector constraint registration calculation. The spatial projection drift error of the geometric feature point cloud of the independent disease area between continuous terrain grids is compensated based on the optical flow trajectory of the de-blurred image set and the pose change of the flight attitude data. After spatial attribute binding of the registered and optimized geometric feature point cloud of the independent disease region with the disease category label, topological reconstruction is performed to output the disease distribution map with spatial coordinates that integrates disease geometric morphology and semantic information.

5. The method for autonomous perception and inspection of highway slope defects based on unmanned aerial vehicles (UAVs) according to claim 4, characterized in that, After spatial attribute binding of the registered and optimized geometric feature point cloud of the independent disease region with the disease category label, topological reconstruction is performed, including: Based on the semantic association rules of the disease category labels, a spatial topological dependency graph is established among the independent disease regions, and a disease attribute relationship network reflecting the evolution and correlation of geological disasters is constructed. The distribution characteristics of the directional gradient histogram of the geometric feature point cloud are obtained and structurally coupled with the disease attribute relationship network to verify the consistency parameters of the three-dimensional deformation trend of each independent disease region. Based on the rock strata structure features of the three-dimensional spatial topology map, the disease attribute relationship network is optimized by performing geomechanical constraints to correct the pseudo-disease point cloud data that does not match the actual terrain. The optimized disease attribute relationship network is implanted into the disease distribution map to generate a disease risk evolution model with spatial hierarchical visualization capabilities.

6. A UAV-based autonomous sensing and inspection system for highway slope defects, characterized in that: The system employs the UAV-based autonomous perception and inspection method for highway slope defects as described in claim 1, wherein the system comprises: The information acquisition module uses drones to collect time-series image streams of the road area to be inspected and flight attitude data at the corresponding acquisition time. The spatial topology module constructs an updatable 3D spatial topology map based on temporal image streams and flight attitude data; The flight control module performs closed-loop control of the UAV's flight attitude based on the real-time positioning deviation of the three-dimensional spatial topology map. The image processing module dynamically adjusts the UAV's flight altitude and image acquisition frequency based on a preset ground resolution threshold, generates a set of de-blurred images, and performs illumination invariance enhancement processing on the de-blurred image set. The disease labeling module obtains disease category labels and structural masks from the images of the motion-blurred image set after illumination invariance enhancement processing based on the topological segmentation method that integrates semantic features. The map generation module maps the structural mask onto a three-dimensional spatial topological map, outputting a disease distribution map with spatial coordinates.

7. The UAV-based autonomous sensing and inspection system for highway slope defects according to claim 6, characterized in that, The disease marking module includes: The feature convolution unit performs multi-level convolutional feature extraction on the motion-blurred image set after illumination invariance enhancement, generating a semantic feature map set including local texture and global structure. The initial segmentation unit inputs the semantic feature map into the attention-guided feature fusion network, which adaptively enhances the edge response of the diseased area and generates an initial segmentation mask with topological connectivity. Geometric topology units, based on the continuity constraints of road slope structures, perform topological relationship optimization on the initial segmentation mask to eliminate isolated noise segmentation regions and repair fractured geometric boundaries; The label matching unit marks the corresponding geological hazard type label on the optimized initial segmentation mask, and simultaneously outputs the hazard category label and the structural mask that conforms to the continuity of the road slope structure.

Images