A deep learning-based target-free slope visual deformation monitoring method and system

Abstract

The present application provides a kind of based on deep learning's no target slope visual deformation monitoring method and system, the present application aims at solving the technical problem that existing slope monitoring technique relies on artificial target, feature matching is unstable under complex natural environment, cannot distinguish camera microvibration and real deformation of slope.The present application extracts slope natural texture feature point using SuperPoint network, carries out feature matching through LightGlue network and is screened initially using confidence, estimates homography matrix using RANSAC algorithm to compensate camera global motion, calculates feature point re-projection residual as deformation index, realizes risk assessment in combination with regional gridding statistics and multi-time trend analysis.The present application does not need to set artificial target, is strong in robustness under complex scene such as illumination dramatic change, texture sparse repetition, can effectively separate global interference and local deformation, output risk result with engineering interpretation, provides reliable technical support for slope geological disaster early warning.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and engineering monitoring technology, and in particular to a targetless slope visual deformation monitoring method and system based on deep learning. Background Technology

[0002] Slope engineering is a common structural form in mining, highway and railway construction, and water conservancy and hydropower projects. Southwest my country is predominantly mountainous, prone to geological disasters. Slope stability monitoring is of significant engineering importance for disaster prevention and mitigation. Traditional monitoring technologies include: 1. Manual inspection: highly subjective, dangerous, and difficult to quantify and provide real-time data; 2. Sensor monitoring (displacement gauges, stress gauges): point-based monitoring, complex deployment, and difficulty in ensuring data continuity under adverse weather conditions; 3. GNSS positioning: high equipment cost, satellite signals are easily blocked in mountainous areas / deep mines; 4. InSAR technology: long satellite revisit cycles, insufficient temporal resolution, and difficulty in meeting the need for instantaneous early warning before landslides; 5. GB-SAR / UAV photogrammetry: expensive equipment and complex data processing.

[0003] In recent years, computer vision technology has provided new solutions for slope monitoring due to its advantages such as non-contact measurement, flexible deployment, and rich information. Existing visual monitoring technologies mainly include the following categories: The visual monitoring scheme based on artificial targets involves installing artificial targets on the slope surface and using image processing algorithms to identify changes in target position to calculate displacement. However, this method requires the pre-installation of artificial markers on the slope surface. For steep, dangerous, or inaccessible slopes, target placement is difficult or even impossible. In the field, targets are easily damaged or detached due to wind and rain erosion, rockfalls, and vegetation obstruction, leading to monitoring interruptions. Furthermore, the limited number of targets only allows for discrete point monitoring and cannot reflect the overall slope deformation field.

[0004] Targetless solutions based on traditional feature matching: These solutions employ traditional hand-crafted feature operators such as SIFT, SURF, and ORB for feature extraction and matching. However, these hand-designed feature operators rely on manually preset rules for gradients, grayscale, etc., which significantly degrade their feature description capabilities under drastic lighting conditions. Furthermore, the texture of exposed soil and rock areas on slopes is sparse, making it difficult for traditional operators to detect stable feature points. Regions with repetitive textures are prone to generating numerous false matches, and there is a lack of effective global consistency constraints.

[0005] Deep learning-based feature matching schemes, exemplified by SuperGlue, model feature matching as an optimal transport problem on a graph structure. However, the multi-layered iteration of graph neural networks leads to slow inference speeds, limiting their application in multi-time-series monitoring scenarios where efficiency is critical; they only complete feature matching without providing a complete technical link from matching results to deformation assessment; and they are not optimized for specific needs such as camera micro-vibration and long-term time-series analysis in slope monitoring.

[0006] A targetless monitoring scheme based on digital image correlation (DIC) is proposed: This scheme uses digital image correlation methods to process natural textures. However, the DIC algorithm is extremely sensitive to changes in illumination, and drastic fluctuations in illumination in the field can cause correlation calculations to fail. Furthermore, the computational cost of correlation searches at the integer or sub-pixel level is too high, making it difficult to meet the needs of real-time monitoring.

[0007] The common drawbacks of the above-mentioned schemes include: failure to effectively distinguish between the overall camera offset (caused by wind and vibration) and the actual local deformation of the slope; analysis results remaining at the point level or whole map level, failing to achieve regional risk classification with engineering interpretability; and failure to establish a residual change sequence of multi-temporal data, which cannot reflect the dynamic evolution trend of the slope. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a targetless slope visual deformation monitoring method and system based on deep learning. Through deep learning feature matching and homography model global compensation, the invention effectively separates camera micro-vibration interference from the actual slope deformation, and realizes high-precision and robust quantitative assessment of slope deformation and dynamic risk warning in complex natural environments.

[0009] According to one objective of the present invention, the present invention provides a targetless slope visual deformation monitoring method based on deep learning, comprising the following steps: S1. Region of Interest (ROI) Selection: In the reference image, a rectangular cropping region is set according to the spatial distribution range of the slope. The main area of ​​the slope is cropped to obtain the ROI image. S2. Deep learning feature extraction: The SuperPoint network is used to detect and describe feature points in the ROI image, and outputs the coordinates of the feature points and their corresponding 256-dimensional descriptors. S3. Feature matching and confidence filtering: The LightGlue network is used to match feature points of the reference image and the target image, output matching point pairs with confidence, and set a confidence threshold to retain only the matching relationship with a confidence higher than the threshold. S4. Coordinate unification: Based on the position of the ROI in the original image, the coordinates of the feature points within the ROI are translated and mapped to the coordinate system of the original image; S5. Homography Model Estimation: The RANSAC algorithm is used to robustly estimate the selected matching points and calculate the homography matrix H between the reference image and the target image. S6. Residual Calculation: Use the homography matrix H to map the coordinates of the interior feature points in the reference image, predict their theoretical positions in the target image, and calculate the Euclidean distance between the actual matching position and the theoretical predicted position as the residual. S7. Residual-based deformation analysis: The residuals are analyzed from three dimensions: overall statistics, outlier identification, and regional analysis to identify potential risk areas; S8. Time-series variation analysis: Integrate the regional average residual results at different time points to construct a sequence of residual changes in each region over time, and analyze the dynamic evolution trend of slope deformation.

[0010] Further, in step S2, the SuperPoint network adopts an encoder-decoder architecture. The encoder is a VGG-style convolution stack, which downsamples the input image by 8 times through three 2×2 max pooling operations to extract multi-scale shared features. On top of these shared features, the network is divided into two parallel decoders: one is a feature point decoder, which outputs a pixel-wise interest point probability map, which is then processed by softmax to remove auxiliary channels and rearranged to obtain the feature point positions; the other is a feature point descriptor decoder, which outputs a semi-dense 256-dimensional descriptor grid, which is then upsampled to the original image size through bicubic interpolation and L2 normalized to obtain the unit descriptor vector corresponding to each pixel.

[0011] Furthermore, in step S2, during inference, the input images are uniformly converted into grayscale images, and a maximum of 2048 feature points with the highest confidence are retained per frame.

[0012] Furthermore, in step S3, the confidence threshold is set to 0.6.

[0013] Furthermore, in step S5, the reprojection threshold of the RANSAC algorithm is set to 3.0 pixels, the confidence level is set to 0.99, and the number of iterations is dynamically adjusted according to the initial matching inlier rate.

[0014] Further, in step S7, the overall statistics include calculating the mean μ and standard deviation σ of all in-point residuals; outlier identification adopts the 3σ criterion, and when the residual exceeds μ+3σ, it is judged as an outlier; the regional analysis includes dividing the ROI into regular sub-regions, calculating the average value of the feature point residuals in each sub-region as the regional deformation index, setting an adaptive risk threshold T=μ+2σ, and marking it as a potential risk region when the average residual of the region exceeds the threshold.

[0015] Furthermore, in step S8, the slope condition is determined by observing the trend of the residual sequence: continuous increase of residual indicates that the slope is undergoing a gradual deformation process; abrupt change in residual indicates that it may be a short-term disturbance caused by external factors; and stable residual indicates that the slope is generally stable.

[0016] According to another objective of the present invention, the present invention provides a targetless slope visual deformation monitoring system based on deep learning, comprising: The image input module is used to read slope image data acquired at different times; The ROI selection module is used to identify the region of interest in a reference image and reduce background interference. The feature extraction and matching module is used to extract image features and establish feature point matching relationships across time phases. The geometric modeling module is used to estimate geometric transformation models between images based on matching points; The residual analysis module is used to calculate the reprojection error of the matching points and perform statistical analysis. The results output module is used to generate residual distribution maps, regional statistical results, risk assessment maps, and time series statistical maps of regional average residuals.

[0017] Furthermore, the feature extraction and matching module includes: The feature extraction unit uses the SuperPoint network to detect and describe feature points in the ROI image, and outputs the coordinates of the feature points and their corresponding 256-dimensional descriptors. The feature matching unit uses the LightGlue network to match feature points between the reference image and the target image, and outputs matching point pairs with confidence scores. The confidence level filtering unit is used to set a confidence level threshold, retaining only matches with a confidence level higher than the threshold.

[0018] Furthermore, the residual analysis module includes: The residual calculation unit is used to perform coordinate mapping on the interior feature points in the reference image using the homography matrix H, predict their theoretical positions in the target image, and calculate the Euclidean distance between the actual matching position and the theoretical predicted position as the residual. The overall statistical unit is used to calculate the mean μ and standard deviation σ of all in-point residuals; Anomaly identification unit, used to identify anomalies using the 3σ criterion; The region analysis unit is used to divide the ROI into regular sub-regions, calculate the average residual of feature points in each sub-region as the region deformation index, set an adaptive risk threshold T=μ+2σ, and mark the region as a potential risk region when the average residual of the region exceeds the threshold. The time series analysis unit is used to integrate the regional average residual results at different time points, construct the sequence of residual changes in each region over time, and analyze the dynamic evolution trend of slope deformation.

[0019] The technical solution of this invention eliminates the need for manual target deployment, directly utilizing the natural texture of the slope to achieve automated monitoring. This solves the pain points of difficult target deployment in steep and dangerous areas and the easy damage and detachment of targets in the field environment. It employs a combination of SuperPoint and LightGlue to achieve stable feature matching in complex environments, and combines confidence-based initial screening with RANSAC geometric fine screening to significantly reduce the mismatch rate in scenes with drastic changes in lighting and sparse and repetitive textures. It effectively separates the global migration of camera micro-vibrations from the local real deformation of the slope through a homography model. Using reprojection residuals as a deformation index, it combines global consistency and anti-interference capabilities. With the addition of regional statistics and time series analysis, it can output risk assessment results with engineering interpretability. Attached Figure Description

[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the overall process of the deep learning-based targetless slope visual deformation monitoring system according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the decoder structure of the SuperPoint network according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the LightGlue network architecture according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the implementation process of homography matrix estimation in an embodiment of the present invention; Figure 5 This is a visualization result of the inner and outer points of the slope monitoring image in an embodiment of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0023] This invention aims to solve the following technical problems existing in the prior art: Traditional monitoring methods are costly to deploy and have limited coverage. GNSS is also susceptible to signal blockage in scenarios such as canyons. Existing visual monitoring methods rely on manual targets, which are easily damaged or detached in complex outdoor environments. Furthermore, target deployment is difficult or even impossible for large-scale, inaccessible slope scenarios. Feature extraction and matching are unstable in complex natural environments, and the discrimination ability of traditional manual feature operators degrades significantly under conditions such as drastic changes in illumination, sparse texture, and repetitive texture. In targetless mode, deformation quantization lacks physical reference. The feature point distribution is discrete and lacks physical reference benchmarks, making it difficult to convert pixel-level changes into deformation evaluation indicators with engineering significance. Camera micro-vibration interference and local deformation are difficult to distinguish. The slight displacement caused by wind and foundation micro-vibration of fixed monitoring equipment and the real local deformation of slope are both shown as pixel displacement in the image, which is difficult to effectively distinguish using traditional direct displacement analysis methods.

[0024] Example 1 like Figure 1 As shown in the figure, this embodiment presents a targetless slope visual deformation monitoring method based on deep learning. The overall technical solution includes the following core steps: Step 1: Region of Interest (ROI) Selection In the reference image, based on the spatial distribution of the slope within the image, the boundary parameters [y_min, y_max, x_min, x_max] of the rectangular cropping region are manually set to crop the main area of ​​the slope. This processing effectively suppresses interference from irrelevant backgrounds such as the sky and distant scenery, reduces computational complexity, and ensures that images from different time periods can be compared and analyzed within the same spatial range.

[0025] After cropping, the image is input into the subsequent model while maintaining its original resolution to preserve image detail. In subsequent processing, the feature point coordinates are uniformly mapped back to the original image coordinate system, ensuring data consistency across different stages.

[0026] Step 2: Deep Learning Feature Extraction Based on SuperPoint The SuperPoint network is used for feature point detection and description of ROI images. SuperPoint is a fully convolutional neural network based on self-supervised learning that can simultaneously output the location of feature points and their corresponding 256-dimensional descriptors in a single forward propagation, overcoming the limitation of traditional methods where detection and description are performed step by step.

[0027] like Figure 2As shown, SuperPoint employs an encoder-decoder architecture. The encoder is a VGG-style stacked convolution, downsampling the input image by 8 times through three 2×2 max pooling operations to extract multi-scale shared features. Above these shared features, the network consists of two parallel decoders: one is a feature point decoder, outputting a pixel-wise interest point probability map with a channel dimension of 65, corresponding to 64 pixel positions in an 8×8 local region plus a feature-free channel, which is then softmaxed to remove auxiliary channels and rearranged; the other is a feature point descriptor decoder, outputting a semi-dense 256-dimensional descriptor grid, which is then upsampled to the original image size through bicubic interpolation and L2 normalized to obtain the unit descriptor vector corresponding to each pixel. Both decoders avoid using learnable upsampling layers, effectively controlling computational overhead while maintaining accuracy.

[0028] During inference, the input images are uniformly converted to grayscale, and the highest 2048 feature points with the highest confidence are retained in each frame.

[0029] Step 3: Feature matching and confidence screening based on LightGlue In the feature matching stage, after obtaining the feature points and descriptors extracted by SuperPoint, the system uses LightGlue for feature matching between images. LightGlue is a Transformer-based deep neural network used for joint matching and outlier removal of sparse local features. LightGlue considers all feature points of two images simultaneously and uses an attention mechanism to infer correspondences in the global context, maintaining high matching robustness even in scenes with drastic changes in lighting and viewpoint.

[0030] like Figure 3 As shown, LightGlue receives the feature point coordinates and 256-dimensional descriptors output by SuperPoint, and outputs matching point pairs with confidence scores. A filtering mechanism based on matching confidence is introduced, setting a confidence threshold of 0.6, retaining only matches with confidence scores higher than the threshold, thus filtering out low-quality matches at the source.

[0031] Step 4: Coordinate unification Since feature extraction and matching are performed within the Region of Interest (ROI), its coordinates are referenced to the local region. The coordinates are translated and mapped based on the ROI's position in the original image to restore the original image coordinate system, ensuring that subsequent geometric modeling and residual calculations are performed within a unified coordinate system.

[0032] Step 5: Homography model estimation based on RANSAC The RANSAC algorithm is used to robustly estimate the selected matching points, and the homography matrix H between the reference image and the target image is calculated.

[0033] To strike a balance between monitoring accuracy and computational efficiency, the system sets `ransacReprojThreshold` to 3.0 pixels. This value is chosen considering the system's goal of achieving millimeter-level accuracy; a tolerance of 3 pixels effectively absorbs sub-pixel quantization noise generated by image digitization while remaining sensitive to local anomalies exceeding the normal range. The default confidence level is set to 0.99, and the number of iterations is dynamically adjusted based on the initial matching inlier rate.

[0034] like Figure 4 As shown, the system calculates the homography matrix H and the interior point mask Mask using the cv2.findHomography function. Then, it uses the obtained H matrix to perform coordinate mapping on the feature point P(x,y) in the reference image to predict its theoretical position P'(x',y') in the monitoring image. The specific transformation is shown in Formula 1: (1); In the formula, H is a 3×3 homography matrix used to describe the overall projection relationship between images. By estimating this matrix, coordinate mapping from the reference image to the target image can be achieved.

[0035] After obtaining the homography model, it can be used to predict the position of feature points in the reference image. The system calculates the Euclidean distance between the actual matching point P1 and the theoretically predicted point P' as the model residual, which reflects the degree to which the feature point deviates from the overall motion model. The system records the dx_res, dy_res, and total residual distance res_dist for each point. The res_dist is calculated as follows (Formula 2): (2); The smaller the residual, the better the point conforms to the overall motion model; conversely, a larger residual indicates a significant local shift at that point. To visually evaluate the RANSAC filtering effect, the processing results are visualized, such as... Figure 5 As shown in the figure, the green dots represent the interior points that conform to the overall motion model. They are mostly distributed in stable areas such as lattice beams and bedrock, and these points constitute the reference coordinate system for monitoring. The red dots represent the exterior points that have been removed, mainly concentrated in vegetation-covered areas and areas with low matching quality.

[0036] It should be noted that, ideally, if the slope remains rigid and does not undergo local deformation, most feature points should be well described by a single homography model, and the residual distribution should be relatively concentrated and the values ​​small. However, in practical engineering applications, due to factors such as local structural changes, minor displacements, or lighting disturbances, some feature points may deviate from the overall movement trend, exhibiting larger residual values. Therefore, the residuals not only reflect the matching error but also, to some extent, reflect the variation characteristics of local areas.

[0037] This embodiment introduces the RANSAC algorithm to construct a robust homography model and calculates feature point residuals based on this model, thus achieving an effective characterization of local changes between images. This process not only provides a data foundation for subsequent residual statistical analysis but also lays a crucial foundation for the quantitative assessment of slope deformation.

[0038] Step Six: Residual Calculation Using the estimated homography matrix H, coordinate mapping is performed on the interior feature points P(x,y) in the reference image to predict their theoretical positions P'(x',y') in the target image. The Euclidean distance between the actual matching position and the theoretical predicted position is calculated as the residual.

[0039] The residual reflects the degree to which a feature point deviates from the overall motion model: the smaller the residual, the more the point conforms to the overall motion model (is located in a stable region); the larger the residual, the more obvious the local offset (possibly due to deformation) exists at the point.

[0040] Step 7: Deformation Analysis Based on Residues The analytical methodology is constructed from three dimensions, focusing on the residual index: Overall statistics: Calculate the mean μ and standard deviation σ of all in-point residuals. The mean represents the overall level of deviation, and the standard deviation characterizes the degree of dispersion.

[0041] Outlier identification: The 3σ criterion is used. When the residual exceeds μ+3σ, it is identified as an outlier (normal probability <0.3%).

[0042] Region analysis: The ROI is divided into regular sub-regions (e.g., a 2×3 grid), and the average residual of feature points within each sub-region is calculated as the region deformation index. An adaptive risk threshold T = μ + 2σ is set; when the average residual of a region exceeds the threshold, it is marked as a potential risk region.

[0043] Step 8: Time Series Variation Analysis The regional average residual results at different time points were integrated to construct a sequence of residual changes over time for each region. By observing the sequence trends: continuous increase in residuals indicates a gradual deformation process in the slope; abrupt changes in residuals indicate possible short-term disturbances caused by external factors; and stable residuals indicate overall slope stability.

[0044] In practical applications, the residual distribution of different feature points often exhibits a certain degree of non-uniformity. This non-uniformity stems not only from differences in the local structure of the slope but may also be influenced by the density of feature point distribution. Therefore, when conducting regional residual analysis, it is necessary to pay attention to the impact of the number of feature points on the statistical results. When the number of feature points in a certain region is small, its average residual may be significantly affected by individual outliers, thereby reducing the stability of the analysis results. In further research, a minimum number of feature points constraint can be introduced into the regional analysis process, that is, only regions that meet a certain number of feature points are effectively statistically analyzed, thereby improving the reliability of the analysis results.

[0045] From an engineering application perspective, the spatial continuity of residuals is also of great significance. When high residual points exhibit a certain clustering trend in space, they often correspond to localized areas of change in the actual structure, while randomly distributed high residual points are more likely caused by noise or matching errors. Therefore, when conducting risk assessments, both the numerical magnitude and spatial distribution characteristics of the residuals should be considered comprehensively, and the continuity of residuals within the region should be analyzed to improve the accuracy of deformation identification.

[0046] In multi-time series analysis, the rate of change of residuals can be further examined. Compared to the residual value at a single moment, the rate of change of residuals can more sensitively reflect the evolution trend of slope condition. When the residuals in a certain area increase rapidly in a short period of time, it often indicates that there may be a risk of accelerated deformation in that area. By introducing the rate of change of residuals as an auxiliary indicator, richer information can be provided for slope stability assessment.

[0047] From an overall methodological perspective, residual-based deformation analysis achieves a mapping from geometric deviations to structural changes. On the one hand, residual statistics can quantitatively describe the overall state of the slope; on the other hand, this paper, through regional division and time-series analysis, can achieve a deeper characterization of local changes and dynamic trends. Compared with traditional methods that rely solely on displacement, this method pays more attention to the deviation of local areas from the overall motion, and has certain advantages in identifying minute deformations.

[0048] Example 2: Targetless Slope Visual Deformation Monitoring System Based on Deep Learning This invention also provides a targetless slope visual deformation monitoring system based on deep learning, comprising: Image input module: Responsible for reading slope image data acquired at different times, and supports input of multiple image formats.

[0049] ROI selection module: Determines the region of interest in the reference image, reduces background interference, and supports both manual interactive rectangular selection and parametric input.

[0050] Feature extraction and matching module: Extracts image features and establishes feature point matching relationships across time phases, including feature extraction unit, feature matching unit and confidence screening unit.

[0051] Geometric modeling module: Based on the geometric transformation model between images estimated by matching points, robust estimation is performed using the RANSAC algorithm, and the homography matrix and interior point mask are output.

[0052] The residual analysis module calculates the reprojection error of the matching points and performs statistical analysis, including a residual calculation unit, an overall statistical unit, an outlier identification unit, a region analysis unit, and a time series analysis unit.

[0053] Results output module: Generates residual distribution map, regional statistical results, risk assessment map and regional average residual time series statistical map, supporting export in multiple formats and visualization display.

[0054] Compared with the prior art, the present invention has the following beneficial effects: This invention eliminates the need for manual target deployment: it directly utilizes the natural texture of the slope surface (stones, cracks, vegetation edges, etc.) to achieve automated monitoring, solving the monitoring challenges of steep, dangerous, or inaccessible slopes, avoiding the vulnerability of targets, and reducing deployment and maintenance costs.

[0055] This invention exhibits strong robustness in matching under complex environments: by combining self-supervised deep learning feature extraction (SuperPoint) with lightweight graph neural network matching (LightGlue), it can maintain high matching accuracy and stability even under complex natural conditions such as drastic changes in lighting, sparse textures, and repetitive textures.

[0056] This invention can effectively separate global motion from local deformation: by estimating the homography matrix using the RANSAC algorithm to characterize the perspective transformation relationship between the camera and the slope, it effectively compensates for the overall offset caused by camera micro-vibration, solving the problem that traditional methods cannot distinguish between camera offset and real deformation.

[0057] The deformation characterization of this invention is more reliable: it uses the reprojection error of feature points under the constraints of the global geometric model as the core deformation index. Compared with the traditional direct displacement analysis method, it has stronger global consistency and anti-interference ability, and its sensitivity to individual mismatches is greatly reduced.

[0058] This invention is highly practical for engineering applications: it realizes residual statistics and adaptive risk threshold determination in gridded areas, and the results are presented in the form of "risk area / stable area". Combined with time series trend charts, it is more in line with the cognitive habits of engineering monitoring personnel and facilitates rapid decision-making.

[0059] This invention boasts high computational efficiency: it employs the LightGlue network and introduces an adaptive depth and width mechanism to dynamically adjust the computational load based on the matching difficulty, significantly reducing inference latency while ensuring matching accuracy, making it more suitable for long-term monitoring scenarios with multiple time series.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A targetless slope visual deformation monitoring method based on deep learning, characterized in that, Includes the following steps: S1. Region of Interest (ROI) Selection: In the reference image, a rectangular cropping region is set according to the spatial distribution range of the slope. The main area of ​​the slope is cropped to obtain the ROI image. S2. Deep learning feature extraction: The SuperPoint network is used to detect and describe feature points in the ROI image, and outputs the coordinates of the feature points and their corresponding 256-dimensional descriptors. S3. Feature matching and confidence filtering: The LightGlue network is used to match feature points of the reference image and the target image, output matching point pairs with confidence, and set a confidence threshold to retain only the matching relationship with a confidence higher than the threshold. S4. Coordinate unification: Based on the position of the ROI in the original image, the coordinates of the feature points within the ROI are translated and mapped to the coordinate system of the original image; S5. Homography Model Estimation: The RANSAC algorithm is used to robustly estimate the selected matching points and calculate the homography matrix H between the reference image and the target image. S6. Residual Calculation: Use the homography matrix H to map the coordinates of the interior feature points in the reference image, predict their theoretical positions in the target image, and calculate the Euclidean distance between the actual matching position and the theoretical predicted position as the residual. S7. Residual-based deformation analysis: The residuals are analyzed from three dimensions: overall statistics, outlier identification, and regional analysis to identify potential risk areas; S8. Time-series variation analysis: Integrate the regional average residual results at different time points to construct a sequence of residual changes in each region over time, and analyze the dynamic evolution trend of slope deformation.

2. The deep learning-based visual deformation monitoring method for targetless slopes according to claim 1, characterized in that, In step S2, the SuperPoint network adopts an encoder-decoder architecture. The encoder is a VGG-style convolution stack, which downsamples the input image by 8 times through three 2×2 max pooling operations to extract multi-scale shared features. On top of these shared features, the network consists of two parallel decoders: one is a feature point decoder, which outputs a pixel-wise interest point probability map, which is then processed by softmax to remove auxiliary channels and rearranged to obtain the feature point positions; the other is a feature point descriptor decoder, which outputs a semi-dense 256-dimensional descriptor grid, which is then upsampled to the original image size through bicubic interpolation and L2 normalized to obtain the unit descriptor vector corresponding to each pixel.

3. The deep learning-based visual deformation monitoring method for targetless slopes according to claim 1, characterized in that, In step S2, during inference, the input images are uniformly converted to grayscale images, and a maximum of 2048 feature points with the highest confidence are retained per frame.

4. The deep learning-based visual deformation monitoring method for targetless slopes according to claim 1, characterized in that, In step S3, the confidence threshold is set to 0.

6.

5. The deep learning-based visual deformation monitoring method for targetless slopes according to claim 1, characterized in that, In step S5, the reprojection threshold of the RANSAC algorithm is set to 3.0 pixels, the confidence level is set to 0.99, and the number of iterations is dynamically adjusted according to the initial matching inlier rate.

6. The deep learning-based visual deformation monitoring method for targetless slopes according to claim 1, characterized in that, In step S7, the overall statistics include calculating the mean μ and standard deviation σ of all in-point residuals; outlier identification uses the 3σ criterion, and when the residual exceeds μ+3σ, it is judged as an outlier; the regional analysis includes dividing the ROI into regular sub-regions, calculating the average value of the feature point residuals in each sub-region as the regional deformation index, setting an adaptive risk threshold T=μ+2σ, and marking a potential risk region when the average residual of the region exceeds the threshold.

7. The deep learning-based visual deformation monitoring method for targetless slopes according to claim 1, characterized in that, In step S8, the slope condition is determined by observing the trend of the residual sequence: continuous increase of residual indicates that the slope is undergoing a gradual deformation process; abrupt change in residual indicates that it may be a short-term disturbance caused by external factors; and stable residual indicates that the slope is generally stable.

8. A targetless slope visual deformation monitoring system based on deep learning, characterized in that, include: The image input module is used to read slope image data acquired at different times; The ROI selection module is used to identify the region of interest in a reference image and reduce background interference. The feature extraction and matching module is used to extract image features and establish feature point matching relationships across time phases. The geometric modeling module is used to estimate geometric transformation models between images based on matching points; The residual analysis module is used to calculate the reprojection error of the matching points and perform statistical analysis. The results output module is used to generate residual distribution maps, regional statistical results, risk assessment maps, and time series statistical maps of regional average residuals.

9. The deep learning-based targetless slope visual deformation monitoring system according to claim 8, characterized in that, The feature extraction and matching module includes: The feature extraction unit uses the SuperPoint network to detect and describe feature points in the ROI image, and outputs the coordinates of the feature points and their corresponding 256-dimensional descriptors. The feature matching unit uses the LightGlue network to match feature points between the reference image and the target image, and outputs matching point pairs with confidence scores. The confidence level filtering unit is used to set a confidence level threshold, retaining only matches with a confidence level higher than the threshold.

10. The deep learning-based targetless slope visual deformation monitoring system according to claim 8, characterized in that, The residual analysis module includes: The residual calculation unit is used to perform coordinate mapping on the interior feature points in the reference image using the homography matrix H, predict their theoretical positions in the target image, and calculate the Euclidean distance between the actual matching position and the theoretical predicted position as the residual. The overall statistical unit is used to calculate the mean μ and standard deviation σ of all in-point residuals; Anomaly identification unit, used to identify anomalies using the 3σ criterion; The region analysis unit is used to divide the ROI into regular sub-regions, calculate the average residual of feature points in each sub-region as the region deformation index, set an adaptive risk threshold T=μ+2σ, and mark the region as a potential risk region when the average residual of the region exceeds the threshold. The time series analysis unit is used to integrate the regional average residual results at different time points, construct the sequence of residual changes in each region over time, and analyze the dynamic evolution trend of slope deformation.

Images