Slope surface deformation intelligent monitoring method and system based on unmanned aerial vehicle oblique photography

Abstract

This invention discloses an intelligent monitoring method and system for slope surface deformation based on UAV oblique photography, relating to the field of slope surface deformation monitoring and image recognition technology. It includes: S1, acquiring multi-period UAV oblique photography images of the slope from multiple perspectives and generating corresponding multi-period 3D point clouds; S2, preprocessing the multi-period 3D point clouds, including point cloud registration based on rock mass structural surface characteristics, and dividing the point clouds into multiple monitoring sub-regions according to both stratigraphic layering and structural surface segmentation. This invention acquires multi-period slope images and generates 3D point clouds using UAV oblique photography technology, and combines layered variable-angle flight path planning and dual-frequency positioning deviation compensation to solve the problem of limited coverage of traditional monitoring methods. The multi-period point clouds are preprocessed using improved SIFT matching, hybrid denoising, and feature point-guided ICP registration, and monitoring sub-regions are divided according to stratigraphic and structural surfaces and bound with lithological labels, providing high-quality data for accurate analysis.

Description

Technical Field

[0001] This invention relates to the field of slope surface deformation monitoring and image recognition technology, specifically to an intelligent monitoring method and system for slope surface deformation based on UAV oblique photography. Background Technology

[0002] Slope surface deformation monitoring is a crucial aspect of ensuring the safety of engineering projects such as highways, railways, and mines. Traditional monitoring methods, including total stations and GPS technology, rely on manual on-site data collection. While these methods can provide accurate single-point measurements, they struggle to comprehensively cover complex slope surfaces. In recent years, unmanned aerial vehicle (UAV) oblique photogrammetry technology has rapidly developed. By capturing high-resolution images from different angles and generating 3D models, it significantly improves the efficiency and scope of slope surface data acquisition. This technology can reconstruct the fine geometry of slopes, providing a rich data foundation for deformation analysis.

[0003] However, existing methods face a core problem when using dense 3D point clouds generated by UAV oblique photography for slope deformation analysis: the signal intensity of geometric changes caused by natural weathering or localized spalling of the slope surface is highly similar to that of actual potential slip deformation, making it difficult for algorithms based on simple geometric feature comparison to reliably distinguish between them. Such methods are prone to misjudging non-structural surface wear as dangerous deformation, thus triggering false alarms and interfering with engineering decisions. Due to the lack of effective integration of the slope's geomechanical background and deformation physical mechanisms, existing monitoring methods are insufficient in identifying early signs of substantial slippage, making it difficult to accurately capture engineering-significant deformation signals in conventional surface fluctuations, thus limiting the accuracy and timeliness of early warnings. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent monitoring method and system for slope surface deformation based on UAV oblique photography, which solves the problems of limited coverage, easy misjudgment of non-structural wear and structural slippage, high false alarm rate and delayed early warning in existing technologies.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent monitoring method for slope surface deformation based on UAV oblique photography, comprising:

[0006] S1. Acquire multi-period UAV oblique photography images of the slope from multiple perspectives and generate corresponding multi-period 3D point clouds;

[0007] S2. Preprocess the multi-phase three-dimensional point cloud, including point cloud registration based on the characteristics of rock mass structural planes, and divide the point cloud into multiple monitoring sub-regions according to the dual dimensions of stratigraphic layering and structural plane block division, and bind lithological labels to each monitoring sub-region.

[0008] S3. For each monitoring sub-region, extract geometric features, texture features, and temporal variation features to form a three-dimensional visual feature set. The temporal variation features include the spatial offset vector of the point cloud in adjacent time periods. The spatial offset vector is decomposed into the normal component along the slope and the tangential component along the slope.

[0009] S4. Call the geomechanical prior parameters corresponding to the monitored sub-region, perform mechanical rationality verification on the spatial offset vector in the three-dimensional visual feature set, remove feature values ​​that do not conform to the rock mass deformation mechanics logic, and generate feature vectors containing geometric, texture, temporal and mechanical four-dimensional information.

[0010] S5. Input the four-dimensional feature vector into a pre-trained deep learning classification model to identify and distinguish between non-structural wear and structural slip deformation.

[0011] Preferably, acquiring multi-phase UAV oblique photographic images of the slope includes:

[0012] Based on the slope height and gradient, a layered variable angle flight path is planned. The layered variable angle flight path includes a low-altitude close-up flight path and a high-altitude panoramic flight path. The camera tilt angle of the low-altitude close-up flight path is linearly adjusted with the slope.

[0013] Control the drone equipped with a five-lens tilting camera gimbal to fly along a layered variable-angle flight path, and simultaneously trigger the exposure of all lenses to obtain multi-view images;

[0014] Spatial attitude information of images is acquired using a dual-frequency positioning unit, and the positioning deviation is compensated by combining the known coordinates of reference points around the slope.

[0015] Preferably, the preprocessing of multi-phase 3D point clouds includes:

[0016] An improved SIFT feature matching algorithm is used to match multi-view images from UAV oblique photography. The improved SIFT feature matching algorithm adds directional constraints on texture features in rock joints and crack areas, and prioritizes matching feature points in key engineering areas to generate single-phase 3D point clouds. Multi-phase 3D point clouds are composed of multiple sets of single-phase 3D point clouds generated at different monitoring periods.

[0017] A hybrid denoising algorithm combining statistical filtering and region growing is used to denoise single-phase 3D point clouds while preserving continuous point clouds of rock mass structural surfaces.

[0018] The feature point-guided ICP registration optimization algorithm is used to register multi-period 3D point clouds. The feature point-guided ICP registration optimization algorithm uses rock mass structural surface feature points extracted from point clouds at different time periods as registration anchor points.

[0019] Preferably, the extraction of geometric features includes:

[0020] The point cloud curvature is calculated using a weighted neighborhood surface fitting algorithm;

[0021] Calculate the elevation variation coefficient of neighboring points to characterize the local elevation dispersion;

[0022] The neighborhood radius used for calculation is set differently based on the lithological labels of the monitored sub-regions.

[0023] Preferably, the extraction of the texture features includes:

[0024] The local binary mode algorithm is used to encode the texture of the rock mass surface, enhance the texture abrupt change features of crack and spalling areas, and filter out invalid textures in vegetation-covered areas.

[0025] Preferably, the geomechanical prior parameters include: rock mass shear strength and structural plane attitude; the mechanical rationality verification includes determining whether the direction of the spatial offset vector matches the structural plane attitude, and whether the cumulative displacement of the rock mass relative to the reference point within adjacent monitoring periods exceeds the allowable range of the corresponding rock mass shear strength.

[0026] Preferably, the deep learning classification model is an improved CNN-LSTM hybrid classification model, in which the convolutional layer is used to extract spatial features and the long short-term memory layer is used to capture temporal deformation patterns; the deep learning classification model introduces a hard sample mining strategy during the training phase to enhance the ability to distinguish between weak slip and weathering.

[0027] Preferably, after identifying and distinguishing between non-structural wear and structural slip deformation, the method further includes:

[0028] For the monitoring sub-regions where structural slip deformation occurs, a Bayesian-updated dynamic threshold generation algorithm is invoked. Based on the normal probability distribution function of the rock mass mechanical parameters and historical deformation data of the region, the real-time deformation rate threshold and cumulative offset threshold are generated by Bayesian posterior probability calculation.

[0029] Preferably, the method further includes:

[0030] The deformation rate and cumulative offset of the monitored sub-regions where structural slip deformation occurs are continuously tracked and compared with the real-time deformation rate threshold and cumulative offset threshold. Based on the comparison results, a graded early warning operation is performed. The graded early warning operation includes pushing deformation monitoring reports, triggering engineering inspection instructions or initiating emergency response procedures, and outputting the three-dimensional coordinates and deformation vector direction of the deformed area.

[0031] This invention also provides an intelligent monitoring system for slope surface deformation based on UAV oblique photography, comprising:

[0032] The UAV oblique photography data acquisition module is used to acquire multi-phase UAV oblique photography images of the slope and generate corresponding multi-phase 3D point clouds.

[0033] The slope 3D point cloud preprocessing module is used to preprocess multi-phase 3D point clouds, including point cloud registration based on rock mass structural surface characteristics, and dividing the point cloud into multiple monitoring sub-regions according to the dual dimensions of stratigraphic layering and structural surface block division, and binding lithological labels to each monitoring sub-region.

[0034] The multi-feature fusion analysis module is used to extract geometric features, texture features, and temporal change features for each monitoring sub-region to form a three-dimensional visual feature set. It also calls the geomechanical prior parameters corresponding to the monitoring sub-region to perform mechanical rationality verification on the spatial offset vector in the three-dimensional visual feature set, and generates a feature vector containing four-dimensional information of geometry, texture, temporal sequence, and mechanics.

[0035] The deformation intelligent recognition and early warning module is used to input four-dimensional feature vectors into a pre-trained deep learning classification model to identify and distinguish between non-structural wear and structural slip deformation.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] This invention acquires multi-period slope images and generates 3D point clouds using UAV oblique photography technology. Combined with layered variable-angle flight path planning and dual-frequency positioning deviation compensation, it solves the problem of limited coverage in traditional monitoring methods. Improved SIFT matching, hybrid denoising, and feature-point-guided ICP registration preprocessing are applied to the multi-period point clouds. Monitoring sub-regions are divided according to strata and structural planes and lithological labels are bound, providing high-quality data for accurate analysis. By extracting geometric, textural, and temporal features, and combining them with prior geomechanical parameters for mechanical verification of spatial offset vectors, and using an improved CNN-LSTM model, non-structural wear and structural slip deformation are effectively distinguished, solving the problems of easy misjudgment and high false alarm rate in traditional methods. For structural slip areas, a Bayesian dynamic threshold generation algorithm and hierarchical early warning are employed to achieve early deformation identification and timely handling, compensating for the shortcomings of traditional methods in capturing early slip signs and the lag in early warning. Overall, this improves the comprehensiveness, accuracy, and timeliness of slope monitoring, providing reliable protection for engineering safety. Attached Figure Description

[0038] Figure 1 This is a flowchart of the method of the present invention;

[0039] Figure 2 This is a flowchart of the data acquisition and 3D point cloud generation process of the present invention;

[0040] Figure 3 This is a flowchart of the point cloud preprocessing and monitoring sub-region division process of the present invention;

[0041] Figure 4 This is a flowchart of the multi-feature fusion and mechanical verification process of the present invention;

[0042] Figure 5 This is a flowchart of the intelligent deformation recognition and hierarchical early warning system of the present invention;

[0043] Figure 6 This is a system structure diagram of the present invention. Detailed Implementation

[0044] 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.

[0045] Example 1

[0046] Please see Figure 1-5 This embodiment provides an intelligent monitoring method for slope surface deformation based on UAV oblique photography, and provides a detailed description in conjunction with the actual scenario of deformation monitoring of a steep rock slope of a mountain highway. The specific implementation includes the following steps:

[0047] The first step is to acquire multi-period UAV oblique photography images of the slope from multiple perspectives and generate corresponding multi-period 3D point clouds. Before image acquisition, a layered variable-angle flight path needs to be planned based on the slope height and gradient. This path includes a low-altitude close-up layer path and a high-altitude panoramic layer path. The lens tilt angle of the low-altitude close-up layer path is linearly adjusted with the slope gradient. Specifically, with a slope height H of 100m, the flight altitude of the low-altitude close-up layer path is set to H / 5, or 20m. This value is based on the optimal distance for low-altitude close-up monitoring verified in engineering practice. This altitude ensures clear capture of slope details while avoiding collisions between the UAV and the slope. The flight altitude of the high-altitude panoramic layer path is set to 2H, or 200m. This altitude provides panoramic coverage of the entire slope, providing macroscopic data support for subsequent overall deformation analysis. The linear relationship between the lens tilt angle θ and the slope α is set as θ = 0.8α + 2°. The value of this linear coefficient is based on the optimal proportion of the rock slope view coverage. For example, when the local slope α is 60°, the corresponding lens tilt angle θ is 50°, which can maximize the coverage of the rock surface texture and geometry in that area. Subsequently, the UAV equipped with a five-lens tilting camera gimbal is controlled to fly along the planned layered variable angle flight path. The five lenses consist of one vertical lens and four tilting lenses. To ensure that the image overlap meets the requirements of 3D modeling, the interval for synchronously triggering the exposure of all lenses is set to 0.5s. This interval is based on the fact that when the UAV flies at a speed of 5m / s, the UAV flies 2.5m within 0.5s, which can ensure that the horizontal overlap of the images is not less than 80% and the vertical overlap is not less than 70%. During image acquisition, a BDS / GPS dual-frequency positioning unit was used to collect spatial attitude information of the images. This positioning unit has a positioning accuracy of ±2cm, which meets the basic positioning requirements for slope deformation monitoring. Simultaneously, three stable bedrock control points around the slope were selected as reference points, with known coordinates of control point 1 (X=123456.789m, Y=987654.321m, Z=500.123m), control point 2 (X=123457.123m, Y=987653.876m, Z=500.089m), and control point 3 (X=123456.543m, Y=987654.654m, Z=500.156m). A seven-parameter Bursa model was used to compensate for positioning deviations in the spatial attitude information. The formula for the compensation model is:

[0048] ;

[0049] in For the compensated coordinates, Original coordinates For translation parameters, Here, m is the rotation parameter and m is the scale parameter. These parameters are obtained by fitting the coordinates of the reference point using the least squares method. For example, the fitting yields the translation parameter. Rotation parameters With the scale parameter m=1.00002, the error of the spatial pose information of the image after compensation can be controlled within ±3cm, laying the foundation for the accurate generation of subsequent 3D point clouds.

[0050] Next, preprocessing of the multi-phase 3D point clouds is performed, including point cloud registration based on rock mass structural surface features, and dividing the point clouds into multiple monitoring sub-regions according to the dual dimensions of stratigraphic layering and structural surface segmentation, and binding lithological labels to each monitoring sub-region. First, an improved SIFT feature matching algorithm is used to match multi-view images from UAV oblique photography. This algorithm adds directional constraints on texture features in rock mass joints and crack areas, prioritizing the matching of feature points in key engineering areas to generate single-phase 3D point clouds. In the feature point detection stage, feature points with gray-level gradient directions within ±15° of the joint strike are assigned a weight of 1.5 times. This weight coefficient is set based on engineering experience values ​​for joint texture recognition, which strengthens the detection priority of feature points in key areas. Simultaneously, the feature point matching threshold is set to 0.7, based on the minimum matching error requirement of the Euclidean distance of feature points, effectively eliminating incorrectly matched feature points. Key engineering areas include the slope toe, slope top, crack intersections, and the perimeter of the support structure; feature points in these areas are prioritized for matching, improving the modeling accuracy of the 3D point cloud in key locations. After feature matching, a hybrid denoising algorithm combining statistical filtering and region growing is used to denoise the 3D point cloud, preserving continuous point clouds of the rock mass structure. The threshold for statistical filtering is set to 3 times the standard deviation, based on the normal distribution characteristics of point cloud noise. For example, when the standard deviation of point cloud coordinates is 0.01m, the corresponding filtering threshold is 0.03m, which can remove isolated noise points outside this range. The growth threshold for the region growing algorithm is set to 0.02m, based on the flatness requirement of the rock mass structure, i.e., the elevation difference between adjacent points on the structure surface does not exceed 0.02m. This ensures that continuous point clouds of the rock mass structure are not mistakenly removed. The noise ratio of the denoised point cloud can be reduced to below 0.5%, and the integrity of the structure surface is not affected. Subsequently, a feature-point-guided ICP registration optimization algorithm was used to register the multi-period 3D point clouds. This algorithm uses rock mass structural surface feature points extracted from point clouds at different time periods as registration anchor points, with no fewer than 50 anchor points. This number is based on the minimum anchor point requirement for registration accuracy, which can ensure the stability of registration. At the same time, the registration convergence threshold is set to 0.005m, which is based on the accuracy requirements of slope deformation monitoring. After registration, the multi-period point cloud registration error can be controlled within ±5mm, meeting the accuracy requirements of subsequent deformation analysis.After point cloud denoising and registration, the point cloud is divided into multiple monitoring sub-regions based on two dimensions: stratigraphic stratification and structural plane segmentation. Stratigraphic stratification is based on the stratigraphic interface depth obtained from drilling data. For example, the slope has a strongly weathered granite layer at a depth of 0-15m, a moderately weathered granite layer at a depth of 15-40m, and a slightly weathered granite layer below 40m. Structural plane segmentation is based on the joint and fracture orientation obtained from the field geological survey. For example, the area covered by a joint set with a strike of 30°, a dip of 120°, and a dip angle of 60° is divided into one sub-region, and the area covered by a joint set with a strike of 150°, a dip of 240°, and a dip angle of 55° is divided into another sub-region. Each monitoring sub-region is bound to a corresponding lithological label, such as strongly weathered granite, moderately weathered granite, and slightly weathered granite, providing a basis for subsequent differentiated feature extraction.

[0051] Subsequently, for each monitored sub-region, geometric features, texture features, and temporal variation features are extracted to form a 3D visual feature set. The temporal variation features include the spatial offset vectors of point clouds from adjacent time periods, which are decomposed into a normal component along the slope and a tangential component along the slope. In the geometric feature extraction stage, a weighted neighborhood surface fitting algorithm is used to calculate the point cloud curvature, with the following formula:

[0052] ;

[0053] Where k is the curvature value of the point cloud surface. For the weight of the neighboring points, To fit the elevation values ​​of the surface, Here, n represents the actual elevation value, and n is the number of neighboring points. The neighbor radius is set differently based on the lithology label of the monitored sub-region. For example, the neighbor radius for strongly weathered granite layers is set to 0.5m. This value is based on the fact that the surface of strongly weathered rock masses is rough and has large local undulations, requiring a larger neighbor range to cover the undulating features. The neighbor radius for slightly weathered granite layers is set to 0.2m. This value is based on the fact that the surface of slightly weathered rock masses is relatively flat, and the neighbor range can be appropriately reduced to ensure the accuracy of the features. Neighboring point weights... Set according to the inverse ratio of distance, that is , The distance from neighboring points to the target point allows for weighting of nearby points, improving the accuracy of curvature calculation. Simultaneously, the elevation variation coefficient of neighboring points is calculated to characterize local elevation dispersion; the formula is as follows:

[0054] ;

[0055] in This is the coefficient of elevation variation for neighboring points, used to quantify local elevation dispersion. The standard deviation of the elevation of neighboring points. The average elevation of neighboring points is used. This coefficient effectively reflects the elevation dispersion within the monitored sub-region, providing a quantitative indicator for the analysis of rock surface undulations. In the texture feature extraction stage, a local binary mode algorithm is used to encode the rock surface texture, enhancing the abrupt texture changes in crack and spalling areas, and filtering out invalid textures in vegetation-covered areas. The encoding formula is as follows:

[0056] ;

[0057] in For target pixel The local binary pattern encoded value, where P is the number of neighboring pixels. The grayscale value of the center pixel. The grayscale value of the neighboring pixels. For a sign function, when x≥0 When x < 0 P is set to 8, a value chosen based on the neighborhood pixel coverage requirement of the rock texture, which can fully capture the local texture features of the rock surface. Simultaneously, a vegetation grayscale threshold is set; for example, areas with grayscale values ​​between 150 and 200 are identified as vegetation. This threshold is based on the grayscale difference between vegetation and rock, effectively filtering out invalid textures in vegetation-covered areas. Furthermore, the grayscale difference threshold for crack areas is set to 20, based on the grayscale contrast requirement between cracks and rock, which enhances the texture abrupt changes in cracks and spalling areas, improving the effectiveness of texture features. In the temporal variation feature extraction stage, the spatial offset vector of the point cloud between adjacent time periods is first calculated. ,in It is a spatial offset vector. , , These are the components of the vector along the x, y, and z axes of the three-dimensional coordinate system, representing the displacement of the rock mass point cloud in the three coordinate axes within adjacent monitoring periods. It is then decomposed into a normal component along the slope and a tangential component along the slope, where the slope normal vector... The slope tangent vector is calculated by fitting the normal vector of the plane through neighborhood points. The unit vector representing the slope direction, with normal components... tangential component ,in The component of the offset vector along the slope normal. It is a spatial offset vector. The slope normal vector, This represents the component of the offset vector along the tangent of the slope. As the slope tangent vector, this decomposition can accurately obtain the deformation direction characteristics of the rock mass surface, providing data support for subsequent mechanical rationality verification. After integrating the above geometric features, texture features and temporal change features, a three-dimensional visual feature set of each monitoring sub-region can be formed.

[0058] Then, the geomechanical prior parameters corresponding to the monitored sub-region are invoked to perform mechanical rationality verification on the spatial offset vector in the three-dimensional visual feature set. Feature values ​​that do not conform to the rock mass deformation mechanics logic are eliminated, and feature vectors containing four-dimensional information including geometry, texture, time sequence, and mechanics are generated. The geomechanical prior parameters include the rock mass shear strength and the attitude of the structural plane. The rock mass shear strength parameter is determined based on empirical values ​​from the "Engineering Geology Handbook". For example, the internal friction angle φ=25° and cohesion c=0.1MPa for strongly weathered granite, the internal friction angle φ=35° and cohesion c=0.3MPa for moderately weathered granite, and the internal friction angle φ=45° and cohesion c=0.8MPa for slightly weathered granite. The structural plane attitude parameter is obtained based on the field geological survey. For example, the structural plane corresponding to a certain monitored sub-region has a strike of 30°, a dip of 120°, and a dip angle of 60°. The mechanical rationality verification is divided into two levels. The first is to determine whether the direction of the spatial offset vector matches the attitude of the structural plane, that is, to determine the tangential component. The direction of the offset vector is determined by whether it aligns with the dip of the structural plane. If the angle between the offset vector and the dip of the structural plane is greater than 30°, it is deemed to violate the logic of rock mass deformation mechanics. This angle threshold is based on the directional characteristics of rock mass structural plane slippage. Secondly, it is determined whether the offset exceeds the allowable range of the corresponding rock mass shear strength, based on the Mohr-Coulomb criterion. Calculate the maximum shear stress that the rock mass can withstand, where For the shear strength of the rock mass, The normal stress on the rock mass The internal friction angle of the rock mass. The cohesion of the rock mass is used as the basis for conversion into the maximum allowable offset, which is then combined with the deformation modulus of the rock mass. For example, when the shear stress of the rock mass in the monitored sub-region exceeds the maximum shear stress, the corresponding offset exceeds the allowable range and is judged to be inconsistent with mechanical logic. After removing the above two types of feature values ​​that do not conform to mechanical logic, the geometric features (curvature k, elevation variation coefficient) are then considered. Texture features (LBP encoded values), temporal features (normal components) Tangential components The system integrates mechanical features (shear strength matching degree, attitude matching degree) to generate a feature vector containing four-dimensional information. This feature vector integrates multi-dimensional data, which can effectively improve the accuracy of subsequent deformation identification and solve the problem of high false alarm rate caused by traditional methods that rely solely on geometric features.

[0059] The four-dimensional feature vector is then input into a pre-trained deep learning classification model to identify and distinguish between unstructured wear and structural slip deformation. This deep learning classification model is an improved CNN-LSTM hybrid model, where convolutional layers are used to extract spatial features, and long short-term memory layers are used to capture temporal deformation patterns. The model has three convolutional layers: the first layer has 3×3 kernels with 32 kernels, the second layer has 3×3 kernels with 64 kernels, and the third layer has 3×3 kernels with 128 kernels. ReLU is used as the activation function for all layers. This structure is based on the spatial feature extraction requirements of deep learning models, enabling layer-by-layer enhancement of multi-dimensional spatial features. The pooling layer uses 2×2 max pooling, which reduces data dimensionality while preserving key features. Two LSTM layers are used. The model has 64 hidden layer units, which can effectively capture the temporal pattern of rock mass deformation. During the training phase, a hard sample mining strategy is introduced, that is, samples that are misclassified are given twice the weight. This weight coefficient is based on the classification priority of hard samples, which can enhance the model's ability to distinguish between weak slip and weathering. The ratio of training set to test set is set to 7:3. The loss function is cross-entropy loss, the optimizer is Adam, and the learning rate is set to 0.001. This learning rate is based on the empirical value of the optimal learning rate when the model converges. After training, the model can accurately distinguish between non-structural wear and structural slip deformation, and its recognition accuracy can meet the actual needs of engineering monitoring.

[0060] After distinguishing between non-structural wear and structural slip deformation, for the monitoring sub-regions where structural slip deformation occurs, a Bayesian-updated dynamic threshold generation algorithm is invoked. Based on the normal probability distribution function of the rock mass mechanical parameters and historical deformation data of the region, real-time deformation rate thresholds and cumulative offset thresholds are generated through Bayesian posterior probability calculation. The formulas are as follows:

[0061] ;

[0062] in In order to obtain historical deformation data Then, the threshold parameter The posterior probability, For threshold parameters, For historical deformation data, For prior probability, Let be the likelihood function. The initial deformation rate threshold is set to 0.5 mm / d, and the initial cumulative offset threshold is set to 10 mm. These initial thresholds are based on the initial early warning experience value of slope slippage. As the monitoring period increases, the threshold parameters are iteratively updated through Bayesian posterior probability, which can realize the dynamic adjustment of the threshold, making the threshold more consistent with the actual deformation characteristics of the monitored sub-region, and avoiding the problem of early warning lag or false alarm caused by fixed thresholds.

[0063] Finally, the deformation rate and cumulative offset of the monitored sub-regions experiencing structural slip deformation are continuously tracked and compared with real-time thresholds. Based on the comparison results, tiered early warning operations are executed. These tiered early warning operations include pushing deformation monitoring reports, triggering engineering inspection instructions, or initiating emergency response procedures, and outputting the three-dimensional coordinates and deformation vector direction of the deformed area. The criteria for determining tiered early warnings are as follows: when the deformation rate is below the dynamic deformation rate threshold, only a deformation monitoring report is pushed, which includes the three-dimensional coordinates of the deformed area, the deformation vector direction, and a deformation trend analysis; when the deformation rate is between the dynamic deformation rate threshold and 1.5 times the threshold, in addition to pushing a monitoring report, an engineering inspection instruction is triggered simultaneously, and professional personnel are dispatched to the site for on-site verification; when the deformation rate exceeds 1.5 times the dynamic deformation rate threshold, or the cumulative offset exceeds the dynamic cumulative offset threshold, the emergency response procedure is immediately initiated, and early warning information is pushed to relevant management departments, along with the output of the precise three-dimensional coordinates and deformation vector direction of the deformed area, providing data support for emergency response. This step effectively improves the timeliness and accuracy of slope deformation early warning, solves the technical pain point of delayed early warning in traditional monitoring methods, and provides reliable protection for engineering safety.

[0064] Example 2

[0065] Please see Figure 6 The present invention also provides an intelligent monitoring system for slope surface deformation based on UAV oblique photography, used to implement the intelligent monitoring method for slope surface deformation based on UAV oblique photography in Embodiment 1. The system includes:

[0066] The UAV oblique photography data acquisition module's core hardware consists of an industrial-grade UAV equipped with a five-lens oblique photography gimbal and a dual-frequency positioning unit. The gimbal's five lenses include one vertical lens and four oblique lenses, enabling simultaneous acquisition of multi-view images. The dual-frequency positioning unit supports BDS / GPS dual-mode positioning, accurately acquiring the spatial attitude information of the images. The module's software includes a flight path planning submodule and an image preprocessing submodule. The flight path planning submodule can automatically generate layered variable-angle flight paths based on input parameters such as slope height and gradient, while also supporting manual fine-tuning. The image preprocessing submodule can use a seven-parameter Bursa model to compensate for positioning deviations in the spatial attitude information of the images, and then generate multi-stage 3D point clouds through multi-view image matching. The generated 3D point cloud data is transmitted to the slope 3D point cloud preprocessing module through a standardized interface.

[0067] The slope 3D point cloud preprocessing module incorporates an improved SIFT feature matching algorithm, a hybrid denoising algorithm combining statistical filtering and regional growth, and a feature point-guided ICP registration optimization algorithm. These algorithms sequentially perform denoising and multi-stage registration of the 3D point cloud. Simultaneously, the module integrates sub-region division functions for stratigraphic stratification and structural plane segmentation. Based on the input stratigraphic drilling data and structural plane attitude data, it automatically divides the registered 3D point cloud into multiple monitoring sub-regions and binds lithological labels to each sub-region. The preprocessed point cloud data and sub-region division results are then transmitted to the multi-feature fusion analysis module.

[0068] The multi-feature fusion analysis module includes a feature extraction submodule and a mechanical verification submodule. The feature extraction submodule can extract geometric features, texture features, and temporal variation features based on the lithological labels of the monitored sub-regions to form a three-dimensional visual feature set. The mechanical verification submodule has a built-in geomechanical prior parameter database, which can automatically call the rock mass shear strength and structural plane attitude parameters of the corresponding monitored sub-regions to perform mechanical rationality verification on the spatial offset vector in the three-dimensional visual feature set. After removing feature values ​​that do not conform to mechanical logic, a feature vector containing four-dimensional information of geometry, texture, temporality, and mechanics is generated. This feature vector will be transmitted to the deformation intelligent recognition and early warning module.

[0069] The deformation intelligent recognition and early warning module has a built-in pre-trained improved CNN-LSTM hybrid classification model, which can automatically input four-dimensional feature vectors into the model to accurately distinguish between non-structural wear and structural slip deformation. At the same time, the module integrates a Bayesian update dynamic threshold generation algorithm, which can generate dynamic early warning thresholds for the monitored sub-regions where structural slip occurs, and continuously track their deformation rate and cumulative offset. Based on the threshold comparison results, it automatically executes hierarchical early warning operations to complete the push of early warning information and trigger the handling instructions.

[0070] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0071] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A slope surface deformation intelligent monitoring method based on unmanned aerial vehicle oblique photography, characterized in that, include: S1. Acquire multi-period UAV oblique photography images of the slope from multiple perspectives and generate corresponding multi-period 3D point clouds; S2. Preprocess the multi-phase three-dimensional point cloud, including point cloud registration based on the characteristics of rock mass structural planes, and divide the point cloud into multiple monitoring sub-regions according to the dual dimensions of stratigraphic layering and structural plane block division, and bind lithological labels to each monitoring sub-region. S3. For each monitoring sub-region, extract geometric features, texture features, and temporal variation features to form a three-dimensional visual feature set. The temporal variation features include the spatial offset vector of the point cloud in adjacent time periods. The spatial offset vector is decomposed into the normal component along the slope and the tangential component along the slope. S4. Call the geomechanical prior parameters corresponding to the monitored sub-region, perform mechanical rationality verification on the spatial offset vector in the three-dimensional visual feature set, remove feature values ​​that do not conform to the rock mass deformation mechanics logic, and generate feature vectors containing geometric, texture, temporal and mechanical four-dimensional information. S5. Input the four-dimensional feature vector into a pre-trained deep learning classification model to identify and distinguish between non-structural wear and structural slip deformation.

2. The intelligent monitoring method for slope surface deformation based on UAV oblique photography according to claim 1, characterized in that, The acquisition of multi-phase UAV oblique photographic images of the slope includes: Based on the slope height and gradient, a layered variable angle flight path is planned. The layered variable angle flight path includes a low-altitude close-up flight path and a high-altitude panoramic flight path. The camera tilt angle of the low-altitude close-up flight path is linearly adjusted with the slope. Control the drone equipped with a five-lens tilting camera gimbal to fly along a layered variable-angle flight path, and simultaneously trigger the exposure of all lenses to obtain multi-view images; Spatial attitude information of images is acquired using a dual-frequency positioning unit, and the positioning deviation is compensated by combining the known coordinates of the benchmark points around the slope.

3. The intelligent monitoring method for slope surface deformation based on UAV oblique photography according to claim 1, characterized in that, The preprocessing of multi-phase 3D point clouds includes: An improved SIFT feature matching algorithm is used to match multi-view images from UAV oblique photography. The improved SIFT feature matching algorithm adds directional constraints on texture features in rock joints and crack areas, and prioritizes matching feature points in key engineering areas to generate single-phase 3D point clouds. Multi-phase 3D point clouds are composed of multiple sets of single-phase 3D point clouds generated at different monitoring periods. A hybrid denoising algorithm combining statistical filtering and region growing is used to denoise single-phase 3D point clouds while preserving continuous point clouds of rock mass structural surfaces. The feature point-guided ICP registration optimization algorithm is used to register multi-period 3D point clouds. The feature point-guided ICP registration optimization algorithm uses rock mass structural surface feature points extracted from point clouds at different time periods as registration anchor points.

4. The unmanned aerial vehicle photogrammetry-based slope surface deformation intelligent monitoring method according to claim 1, characterized in that, The extraction of geometric features includes: The point cloud curvature is calculated using a weighted neighborhood surface fitting algorithm; Calculate the elevation variation coefficient of neighboring points to characterize the local elevation dispersion; The neighborhood radius used for calculation is set differently based on the lithological labels of the monitored sub-regions.

5. The unmanned aerial vehicle photogrammetry-based slope surface deformation intelligent monitoring method according to claim 1, characterized in that, The extraction of the texture features includes: The local binary mode algorithm is used to encode the texture of the rock mass surface, enhance the texture abrupt change features of crack and spalling areas, and filter out invalid textures in vegetation-covered areas. 6.The UAV photogrammetry-based slope surface deformation intelligent monitoring method according to claim 1, characterized in that, The geomechanical prior parameters include: rock mass shear strength and structural plane attitude; mechanical rationality verification includes determining whether the direction of the spatial offset vector matches the structural plane attitude, and whether the cumulative displacement of the rock mass relative to the reference point within adjacent monitoring periods exceeds the allowable range of the corresponding rock mass shear strength.

7. The unmanned aerial vehicle photogrammetry-based slope surface deformation intelligent monitoring method according to claim 1, characterized in that, The deep learning classification model is an improved CNN-LSTM hybrid classification model, in which convolutional layers are used to extract spatial features and long short-term memory layers are used to capture temporal deformation patterns. The deep learning classification model introduces a hard sample mining strategy during the training phase to enhance the ability to distinguish between weak slip and weathering. 8.The UAV photogrammetry-based slope surface deformation intelligent monitoring method according to claim 1, characterized in that, After identifying and distinguishing between non-structural wear and structural slip deformation, the method further includes: For the monitoring sub-regions where structural slip deformation occurs, a Bayesian-updated dynamic threshold generation algorithm is invoked. Based on the normal probability distribution function of the rock mass mechanical parameters and historical deformation data of the region, the real-time deformation rate threshold and cumulative offset threshold are generated by Bayesian posterior probability calculation. 9.The UAV photogrammetry-based slope surface deformation intelligent monitoring method according to claim 1, characterized in that, The method further includes: The deformation rate and cumulative offset of the monitored sub-regions where structural slip deformation occurs are continuously tracked and compared with the real-time deformation rate threshold and cumulative offset threshold. Based on the comparison results, a graded early warning operation is performed. The graded early warning operation includes pushing deformation monitoring reports, triggering engineering inspection instructions or initiating emergency response procedures, and outputting the three-dimensional coordinates and deformation vector direction of the deformed area.

10. An intelligent monitoring system for slope surface deformation based on UAV oblique photography, applied to the intelligent monitoring method for slope surface deformation based on UAV oblique photography as described in any one of claims 1-9, characterized in that, include: The UAV oblique photography data acquisition module is used to acquire multi-phase UAV oblique photography images of the slope and generate corresponding multi-phase 3D point clouds. The slope 3D point cloud preprocessing module is used to preprocess multi-phase 3D point clouds, including point cloud registration based on rock mass structural surface characteristics, and dividing the point cloud into multiple monitoring sub-regions according to the dual dimensions of stratigraphic layering and structural surface block division, and binding lithological labels to each monitoring sub-region. The multi-feature fusion analysis module is used to extract geometric features, texture features, and temporal change features for each monitoring sub-region to form a three-dimensional visual feature set. It also calls the geomechanical prior parameters corresponding to the monitoring sub-region to perform mechanical rationality verification on the spatial offset vector in the three-dimensional visual feature set, and generates a feature vector containing four-dimensional information of geometry, texture, temporal sequence, and mechanics. The deformation intelligent recognition and early warning module is used to input four-dimensional feature vectors into a pre-trained deep learning classification model to identify and distinguish between non-structural wear and structural slip deformation.

Images