High and steep slope operating state non-contact visual settlement displacement monitoring and operating state evaluation method

Abstract

The application discloses a kind of high and steep slope operating state non-contact vision subsidence displacement monitoring and operating state evaluation method, belong to the technical field of high and steep slope intelligent construction technical field, comprising: obtaining high and steep slope image;Extract the feature point of high and steep slope image, and construct high and steep slope subsidence displacement database;Based on graph convolutional neural network, construct feature matching network model;High and steep slope subsidence displacement database is used to train feature matching network model;The newly obtained high and steep slope image is input into feature matching network model and is matched with features, and the key point pair of two high and steep slope images is extracted;Based on key point pair, calculate high and steep slope multi-measuring point subsidence displacement data;According to high and steep slope multi-measuring point subsidence displacement data, high and steep slope operating state risk assessment and early warning are carried out, and the safety operation of high and steep slope and its adjacent traffic facilities and disaster prevention and mitigation have important significance.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of high and steep slope intelligent construction technology, and particularly relates to a high and steep slope operation state non-contact visual settlement displacement monitoring and operation state evaluation method. BACKGROUND

[0002] With the rapid development of economy, more and more infrastructure constructions such as highways, railways, water conservancy, power and mines are built, and the safety and stability of the slope as an important part of the infrastructure directly affect the safety and benefit of the whole project. High embankment and deep ditch are one of the characteristics of mountainous railways, and with the continuous increase of construction scale, combined with the limitation of mountainous topography and geological conditions, the height of the slope engineering is getting higher and higher, and the slope is getting steeper and steeper. The geological environment of mountainous areas, the composition of slope rock-soil mass, lithology, fissure joint and tectonic inclination are quite different, resulting in a variety of complex slope diseases in mountainous areas, and the safety problem of slope construction is increasingly prominent. The continuous occurrence of safety problems of high and steep slopes in the process of railway construction and operation leads to the phenomenon of interruption of construction and operation, which causes direct or indirect losses difficult to estimate.

[0003] Conventional landslide monitoring uses theodolites, levels, collimators, electromagnetic wave range finders and other instruments for observation, mainly using the methods of collimation line, forward intersection, backward intersection, polar coordinate, angle, level measurement and photoelectric triangulation height measurement to obtain observation results. These devices have many shortcomings in practical application, such as high labor cost, low measurement accuracy and poor real-time performance. At the same time, due to the complex construction site environment, there are certain safety hazards. And most monitoring projects only use a single method for observation, which is limited by economic conditions, instrument configuration, and the difficulty of laying out monitoring benchmarks in landslide areas and poor observation conditions, and the accuracy of the observation results is generally low.

[0004] In recent years, the data obtained by manual monitoring is difficult to effectively manage and analyze, and it is difficult to comprehensively evaluate and predict the state and trend of the slope. With the rapid development of Internet of Things technology, sensor technology, machine vision technology and big data analysis technology, the slope monitoring means has changed from traditional ground measurement to modern intelligent and automatic monitoring system, which can obtain slope deformation information in real time, continuously and accurately. Although new monitoring technologies such as InSAR, three-dimensional laser scanning and unmanned aerial vehicle oblique photography provide new methods for slope deformation and have achieved some results in some projects, there are still problems to be solved, such as low data processing efficiency and insufficient adaptability to extreme environments. SUMMARY

[0005] The present application aims at the above-mentioned deficiencies in the prior art, and provides a high and steep slope operation state non-contact visual settlement displacement monitoring and operation state evaluation method to solve the problems of time-consuming and laborious installation of existing contact type sensing technology, process implementation limitation of traditional non-contact monitoring method and low data processing efficiency of non-contact monitoring method.

[0006] To achieve the above-mentioned purposes, the technical solution adopted by the present application is:

[0007] A high and steep slope operation state non-contact visual settlement displacement monitoring and operation state evaluation method, comprising the following steps:

[0008] S1, obtaining high and steep slope images;

[0009] S2, extracting feature points of the high and steep slope images and constructing a high and steep slope settlement displacement database;

[0010] S3, constructing a feature matching network model based on a graph convolutional neural network;

[0011] S4, training the feature matching network model using the high and steep slope settlement displacement database;

[0012] S5, inputting newly obtained high and steep slope images into the feature matching network model for feature matching and extracting key point pairs of the two high and steep slope images;

[0013] S6, calculating high and steep slope multi-point settlement displacement data based on the key point pairs;

[0014] S7, performing high and steep slope operation state risk evaluation and early warning according to the high and steep slope multi-point settlement displacement data.

[0015] Further, in S2, local features of the two adjacent high and steep slope images are extracted, and based on the local features, key point sets of the two adjacent high and steep slope images are respectively constructed and .

[0016] wherein, represents the position of the key point i ; represents the feature descriptor of the key point i ; represents the confidence of the key point i ; M represents the key point number of the key point i ;

[0017] represents the key point position of the key point j ; represents the key point jFeature descriptors; Indicate key points j Confidence level; N Representing an image j Number of key points;

[0018] The set of key points from multiple images of steep slopes constitutes a database of settlement and displacement of steep slopes.

[0019] Furthermore, step S4 specifically includes the following sub-steps:

[0020] S41. Encode the locations of key points and use them as nodes in the graph;

[0021] S42. Node feature updates are performed alternately using self-attention and cross-attention mechanisms.

[0022] S43. Based on the updated node features, obtain the matching features of the node features;

[0023] S44. Based on the matching features, calculate the matching score matrix of two adjacent high and steep slope images. Then, use dustbin expansion and Sinkhorn regularization to process the matching score matrix in sequence to obtain the matching probability matrix.

[0024] Furthermore, in step S41, the encoder maps the positional encoding to a feature space with the same dimension as the feature descriptor through a layer-by-layer stacked feature projection network. Simultaneously, a learnable scalar correction method is introduced to perform learnable calibration of the confidence level, which is expressed as:

[0025]

[0026] In the formula, This indicates the unified characteristics after fusion; Feature descriptors representing key points; Stacked feature projections representing positional encoding; Indicates the location of key points; It is a learnable scalar; Stacked feature projections representing confidence information; Indicates the confidence level of key points.

[0027] Furthermore, in step S42, node feature updates are performed alternately using self-attention and cross-attention mechanisms. The update process is as follows:

[0028]

[0029] In the formula, Indicates the first Layer time i Feature representation of points; Indicates the first Layer time i Point feature representation Stacked feature projection network First layer Layer information aggregation Stitching process

[0030] The information aggregation is calculated by using an attention mechanism , which is represented as:

[0031] , ,

[0032]

[0033] In the formula, Query vector Learnable linear transformation matrix Learnable linear layer transformation bias parameter Key vector Learnable linear transformation matrix Neighbor node j First layer Feature vector of the layer Learnable linear layer transformation bias parameter Value vector Learnable linear transformation matrix Learnable linear layer transformation bias parameter Node i Edge set corresponding to the neighbor set allowed to interact with it in the current layer message passing Improved attention score

[0034] Further, by adding a learnable two-dimensional bias related to the relative position and a sine-cosine relative position encoding in the self-attention and cross-attention scoring of each layer, the improved attention score is calculated , which is specifically represented as:

[0035] Attention score calculation :

[0036]

[0037]

[0038]

[0039] Attention score-based Before cross attention, a candidate set is constructed for each query point, and then improved attention scores are calculated:

[0040]

[0041] where, is the description of the sub-dimension; represents a learnable relative position bias function; represents the relative position; represents the position vector of the node i ; represents the position vector of the neighbor node j ; represents the stacked feature projection of the relative position bias; is the periodic angle feature, is the Euclidean distance, is the direction angle; represents the construction of the query point candidate set.

[0042] Further, in S43, the matching feature of the node feature is represented as:

[0043]

[0044] where, represents the final matching feature representation of the key point i in image A; represents the learnable linear projection layer weight; represents the fused depth feature of the key point i in image A; represents the final matching feature representation of the key point j in image B; represents the fused depth feature of the key point j in image B; represents the learnable linear projection bias.

[0045] Further, S44 specifically includes:

[0046] The matching score matrix is calculated, which is represented as:

[0047]

[0048] where, represents the matching score matrix;

[0049] The unmatching key points are assigned to each set in the matching score matrix by using the dustbin extension, and the extended score matrix is obtained based on the additional new row and new column, which is represented as:

[0050]

[0051] wherein, denotes the score of the i-th keypoint in image A matching to the dustbin column; i denotes the score of the i-th keypoint in image B matching to the dustbin column; j denotes the default score constant of the dustbin;

[0052] Regularization iteration is performed on the extended score matrix using Sinkhorn regularization to obtain the matching probability matrix .

[0053] Further, in the S4, the loss function of the feature matching network model is represented as:

[0054]

[0055] wherein, denotes the loss of the feature matching network model; denotes the matching probability of the i-th keypoint and the j-th keypoint after Sinkhorn regularization; i denotes the probability of the i-th keypoint matching to the dustbin; j denotes the probability of the i-th keypoint matching to the dustbin; i j , , , respectively denote the set of real matching pairs provided in the training, the set of keypoint indexes all without real corresponding points in image A, and the set of keypoint indexes all without real corresponding points in image B.

[0056] Further, the S6 specifically comprises:

[0057] According to the keypoint pairs extracted on the adjacent two high and steep slope images, the pixel displacement size of the keypoint under the camera canvas coordinates is obtained, the pixel displacement is converted into real displacement, and the high and steep slope settlement displacement data under the unified coordinates is obtained through camera coordinate conversion according to the real displacement.

[0058] The high and steep slope operation state non-contact visual settlement displacement monitoring and operation state evaluation method provided by the application has the following beneficial effects:

[0059] ​​​​​​The present application is based on the high and steep slope multi-point settlement displacement data obtained by the high and steep slope multi-point displacement non-contact monitoring method under complex environment, analyzes the settlement displacement result change of the high and steep slope in the operation stage, and then considers the setting of the slope settlement early warning threshold according to the current specification, the settlement rate, the trend and the stability analysis of the high and steep slope, which is of great significance to the safe operation of the high and steep slope and the adjacent traffic facilities and the disaster prevention and reduction. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 The flowchart of the high and steep slope operation state non-contact visual settlement displacement monitoring and operation state evaluation method of the embodiment of the present application.

[0061] Figure 2 The high and steep slope multi-point settlement displacement monitoring scheme under complex environment of the embodiment of the present application.

[0062] Figure 3 The high and steep slope operation state multi-point settlement displacement of the embodiment of the present application. DETAILED DESCRIPTION

[0063] The specific embodiments of the present application are described below to facilitate the understanding of the present application by those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0064] The high and steep slope operation state non-contact visual settlement displacement monitoring and operation state evaluation method of the embodiment can effectively solve the defects of time-consuming and laborious installation of contact sensing technology, process implementation limitation of traditional non-contact monitoring method and low data processing efficiency of new non-contact monitoring method, and reference Figure 1 , which specifically includes the following contents:

[0065] S1, obtaining a high and steep slope image;

[0066] Reference Figure 2 , the embodiment is configured with a high and steep slope multi-point displacement non-contact monitoring module, which uses a high-definition industrial camera to remotely shoot infrared targets distributed in multiple points of the high and steep slope, and obtains a settlement image sequence of the high and steep slope during operation.

[0067] Among them, based on the monitoring object high and steep slope, according to the monitoring condition of the monitoring point itself, the distance between the monitoring point and the slope, the geometric information of the slope, the dense degree of the monitored point and the monitoring accuracy requirement, the number, direction and camera and lens parameters of the monitoring camera are determined.

[0068] S2, extract feature points of high and steep slope images, and construct a high and steep slope settlement displacement database;

[0069] Specifically, local features of two adjacent high and steep slope images are extracted, and based on the local features, key point sets of the two adjacent high and steep slope images are constructed respectively and key point set B are used for feature matching.

[0070] wherein, denotes the position of the key point i ; denotes the feature descriptor of the key point i ; denotes the confidence of the key point i ; M denotes the key point number of the key point i ;

[0071] denotes the key point position of the key point j ; denotes the feature descriptor of the key point j ; denotes the confidence of the key point j ; N denotes the key point number of the image j ;

[0072] The key point sets of multiple high and steep slope images constitute a high and steep slope settlement displacement database.

[0073] S3, constructing a feature matching network model based on a graph convolutional neural network;

[0074] The purpose of the feature matching network model of the embodiment is to output a soft matching matrix , wherein each element represents the matching probability between the feature point i and the feature point j .

[0075] S4, training the feature matching network model using the high and steep slope settlement displacement database, which specifically includes the following steps:

[0076] S41, encoding the key point position and taking it as a node of a graph;

[0077] ​​Specifically, keypoints are treated as nodes in a graph for feature fusion and matching probability estimation. First, the keypoint locations are encoded. The encoder maps these location codes to a feature space with the same dimension as the feature descriptors using a layer-by-layer stacked feature projection network. Simultaneously, to give high-confidence points greater representation in subsequent attention, a learnable scalar correction method is introduced to adjust the confidence level. The learnable calibration is represented as follows:

[0078]

[0079] In the formula, This indicates the unified characteristics after fusion; Feature descriptors representing key points; Stacked feature projections representing positional encoding; Indicates the location of key points; It is a learnable scalar; Stacked feature projections representing confidence information; Indicates the confidence level of key points

[0080] S42. Node feature updates are performed alternately using self-attention and cross-attention mechanisms. The update process is as follows:

[0081]

[0082] In the formula, Indicates the first Layer time i Feature representation of points; Indicates the first Layer time i Feature representation of points; This represents a stacked feature projection network; Indicates the first Layer information aggregation; This indicates splicing / joining.

[0083] Among them, the attention mechanism is used to compute information aggregation. It is represented as:

[0084] , ,

[0085]

[0086] In the formula, Represents the query vector; Represents a learnable linear transformation matrix; This represents the learnable linear layer transformation bias parameters; Represents the key vector; denote the learnable linear layer transformation matrix; denote the neighbor nodes j In the first layer, the feature vector of the current layer; denote the learnable linear layer transformation bias parameter; denote the value vector; denote the learnable linear layer transformation matrix; denote the learnable linear layer transformation bias parameter; denote the node i allowed to interact with the edge set corresponding to the neighbor set; denote the improved attention score.

[0087] The information aggregation performs intra-image feature aggregation through self-attention, and establishes a matching relationship between two images through cross-attention, and then alternately performs self-attention and cross-attention in each layer. The embodiment is improved by adding a learnable two-dimensional bias related to the relative position and a sine relative position encoding in the self-attention and cross-attention scoring of each layer to improve the set consistency, and then the improved attention score is calculated , which is specifically represented as:

[0088] Calculate the attention score :

[0089]

[0090]

[0091]

[0092] In order to suppress the long-distance false peak and improve the contrast of the real matching pair, before the cross-attention, a candidate set is constructed for each query point according to the "appearance similarity + spatial proximity" , and the logits of the column not in it are set to , that is, the improved attention score is obtained:

[0093]

[0094] In the formula, is a description of the sub-dimension; denote the learnable relative position bias function; denote the relative position; denote the position vector of the node i ; denote the position vector of the neighbor node j ; denote the stacked feature projection of the relative position bias; is a periodic angle feature, is an Euclidean distance, is a direction angle; denotes a query point constructing a candidate set.

[0095] S43, based on the updated node feature, the matching feature of the node feature is obtained, and the final output of the matching feature is represented as:

[0096]

[0097] In the formula, denotes the final matching feature representation of the key point i in image A; denotes the final matching feature representation of the key point i in image A; denotes the final matching feature representation of the key point j in image B; denotes the final matching feature representation of the key point j in image B;W and b are used as learnable items to project the updated features of the network to a new space to better represent the matching of the point pair.

[0098] S44, according to the matching feature, the matching score matrix of the adjacent two high and steep slope images is calculated, and the dustbin extension and Sinkhorn regularization are used to process the matching score matrix in sequence to obtain the matching probability matrix.

[0099] The matching score matrix is calculated, which is represented as:

[0100]

[0101] In the formula, denotes the matching score matrix;

[0102] In order to make the network suppress some key points, the dustbin extension is used to distribute the display score of the unmatched key points to each set in the matching score matrix, and the extended score matrix is obtained based on the extension of the additional new row and new column, which is represented as:

[0103]

[0104] In the formula, denotes the score of the i th key point in image A matching to the dustbin column; denotes the score of the j th key point in image B matching to the dustbin column; denotes the default score constant of the dustbin.

[0105] Applying Sinkhorn regularization to the extended fractional matrix Perform regularization iterations to obtain the matching probability matrix. Each element Representing feature points i and feature points j The probability of matching between them.

[0106] The loss function of the feature matching network model in this embodiment is expressed as follows:

[0107]

[0108] In the formula, This represents the loss of the feature matching network model; This indicates that after Sinkhorn regularization, the th i The key point and the first j The matching probability of each key point; Indicates the first i The probability of a key point matching dustbin; Indicates the first j The probability of a key point matching dustbin; , , Let A, B, C, D, and E represent the set of true matching pairs provided during training, the set of keypoint indices for all points that do not have true corresponding points in image A, and the set of keypoint indices for all points that do not have true corresponding points in image B, respectively.

[0109] The self-training mechanism in this embodiment is as follows:

[0110] A graph matching-based self-training mechanism is adopted as the core method of the training process. Pseudo-labels are generated using network video data, and the label quality and training signal density are improved through the following modules:

[0111] I. Generating Initial High-Quality Pseudo-Labels. First, a feature matching network model is initially trained on an existing limited dataset to obtain a basic matcher. Then, images are sampled at fixed frame intervals from videos of slope scenes downloaded from the internet to construct frame pair combinations, such as:

[0112]

[0113] For each frame, a basic matcher is used to obtain preliminary matching results, which are then fused with the SIFT matcher to obtain a richer initial dataset. Finally, a robust fitting algorithm is used to remove outliers and retain point pairs with good geometric consistency.

[0114] II. Delivering high-quality matching between adjacent frames to distant frame pairs. For each pair of image frames I A and I B , a correlation matrix is constructed, representing the pixel-level matching relationship, if matches and at the same time , matches , it is considered that matches , at this time , in the delivery process, the propagation interval of the initial frame pair is 20 frames, and the interval is doubled after each round of propagation until the number of valid matching points between the image pairs is below the threshold.

[0115] III. Considering that video data may be visually continuous or repetitive, directly using the original frame pair can easily lead to model overfitting, therefore, a reinforcement data enhancement mechanism is introduced, including:

[0116] 1. Randomly cropping, color perturbation, illumination variation, rotation, etc. are performed on the image; 2. Random perspective transformation is introduced to simulate different camera perspectives and projection differences; 3. All enhanced frame pairs retain the original pseudo-label coordinates, and the matching relationship is mapped to the enhanced image through coordinate inverse transformation, and the supervision loss is only calculated on the label point pairs.

[0117] During training, mixed sampling data of standard intra-domain data and pseudo-label data generated by the training mechanism is used, specifically: 1. Each batch randomly contains data from the limited dataset and 2) video pseudo-labels generated by the training mechanism; 2. The maximum likelihood loss is used to optimize all samples, and the backpropagation is only performed on the corresponding point pairs provided by the pseudo-labels; 3. The model is initialized as a pre-trained version of the model.

[0118] S5, input the newly acquired high and steep slope image into the feature matching network model for feature matching, and extract the key point pairs of the two high and steep slope images;

[0119] S6, based on the key point pairs, calculate the high and steep slope multi-point settlement displacement data;

[0120] Referring to Figure 3 , according to the key point pairs extracted on the adjacent two high and steep slope images, the pixel displacement size of the key points in the camera canvas coordinates is obtained, and the pixel displacement is converted into real displacement; according to the real displacement, the camera coordinate conversion is carried out, and the high and steep slope settlement displacement data in the unified coordinates is obtained.

[0121] S7, according to the high and steep slope multi-point settlement displacement data, the operation state risk assessment and early warning of the high and steep slope are carried out;

[0122] Specifically, the high and steep slope operation state regular feedback and evaluation early warning method: based on the monitoring calculation of the high and steep slope multi-measuring point settlement displacement data, the high and steep slope operation stage settlement displacement result change analysis is carried out, and then the current specification combined with the slope settlement early warning threshold setting, the settlement rate, the trend and the stability analysis of the high and steep slope are considered respectively. The above three kinds of data: settlement threshold, settlement rate, trend, that is, the corresponding discrete value obtained from the specification, the first order derivative of the settlement displacement data, the second order derivative of the settlement displacement data, are substituted into the machine learning algorithm, and the classification, fitting and re-clustering according to the current specification are carried out, so as to obtain the evaluation method and grade division of the risk area of the high and steep slope operation state. The high and steep slope operation state evaluation digital platform is built, the displacement state and the evaluation result are regularly uploaded to the high and steep slope operation state evaluation digital platform according to the early warning demand, the data support is provided for the safe operation and disaster prevention and reduction of the high and steep slope and its adjacent traffic facilities, and the safety of the structure and facility is ensured.

[0123] Although the specific embodiments of the application are described in detail with reference to the accompanying drawings, it should not be understood as limiting the protection scope of the patent. Various modifications and variations made by those skilled in the art within the scope described in the claims are still within the protection scope of the patent.

Claims

1. A non-contact visual monitoring method for settlement and displacement and an assessment of operational status of steep slopes, characterized in that, Includes the following steps: S1. Obtain images of steep slopes; S2. Extract feature points from images of steep slopes and construct a database of settlement and displacement of steep slopes; S3. Construct a feature matching network model based on graph convolutional neural networks; S4. Use a high and steep slope settlement displacement database to train a feature matching network model; S5. Input the newly acquired steep slope images into the feature matching network model for feature matching, and extract key point pairs from the two steep slope images; S6. Based on key point pairs, calculate settlement and displacement data of multiple measuring points on steep slopes; S7. Based on the settlement and displacement data of multiple measuring points on steep slopes, conduct risk assessment and early warning of the operational status of steep slopes; S4 specifically includes the following steps: S41. Encode the locations of key points and use them as nodes in the graph; S42. Node feature updates are performed alternately using self-attention and cross-attention mechanisms. S43. Based on the updated node features, obtain the matching features of the node features; S44. Based on the matching features, calculate the matching score matrix of two adjacent high and steep slope images. Then, use dustbin expansion and Sinkhorn regularization to process the matching score matrix in turn to obtain the matching probability matrix. In step S41, the encoder maps the positional encoding to a feature space with the same dimension as the feature descriptor through a layer-by-layer stacked feature projection network. Simultaneously, a learnable scalar correction method is introduced to perform learnable calibration of the confidence level, which is expressed as: In the formula, This indicates the unified characteristics after fusion; Feature descriptors representing key points; Stacked feature projections representing positional encoding; Indicates the location of key points; It is a learnable scalar; Stacked feature projections representing confidence information; Indicates the confidence level of key points; In step S42, node feature updates are performed alternately using self-attention and cross-attention mechanisms. The update process is as follows: In the formula, Indicates the first Layer time i Feature representation of points; Indicates the first Layer time i Feature representation of points; This represents a stacked feature projection network; Indicates the first Layer information aggregation; This indicates splicing / joining. Among them, the attention mechanism is used to compute information aggregation. It is represented as: ， ， In the formula, Represents the query vector; Represents a learnable linear transformation matrix; This represents the learnable linear layer transformation bias parameters; Represents the key vector; Represents a learnable linear transformation matrix; Representing neighboring nodes j In the The feature vector of the layer; This represents the learnable linear layer transformation bias parameters; Represents a value vector; Represents a learnable linear transformation matrix; This represents the learnable linear layer transformation bias parameters; This indicates that in the current layer of message passing, the node i The set of edges corresponding to the set of neighbors that are allowed to interact with it; This indicates the improved attention score; The improved attention scores are obtained by incorporating learnable two-dimensional biases and sine / cosine relative position codes related to relative position into the self-attention and cross-attention scores of each layer. Specifically, it is expressed as: Calculate attention score : Based on attention scoring Before cross-attention, a candidate set is constructed for each query point, and then the improved attention score is calculated: In the formula, To describe the sub-dimension; Represents a learnable relative position bias function; Indicates relative position; Represents a node i The position vector; Representing neighboring nodes j The position vector; Stacked feature projections representing relative positional offsets; It is a periodic angular feature. For Euclidean distance, For direction and angle; This indicates that a candidate set is constructed based on the query points.

2. The method for non-contact visual settlement and displacement monitoring and operational status assessment of steep slopes according to claim 1, characterized in that, In step S2, local features are extracted from two adjacent high and steep slope images. Based on these local features, key point sets are constructed for each of the two adjacent high and steep slope images. and key point set ; in, Indicate key points i Location; Indicate key points i Feature descriptors; Indicate key points i Confidence level; M Indicate key points i Number of key points; Indicate key points j Key points and locations; Indicate key points j Feature descriptors; Indicate key points j Confidence level; N Representing an image j Number of key points; The set of key points from multiple images of steep slopes constitutes a database of settlement and displacement of steep slopes.

3. The method for non-contact visual settlement and displacement monitoring and operational status assessment of steep slopes according to claim 1, characterized in that, In S43, the matching features of node features are represented as follows: In the formula, Represents key points in image A i The final matching feature representation; Represents the learnable linear projection layer weights; Represents key points in image A i The fused depth features; Represents key points in image B j The final matching feature representation; Represents key points in image B j The fused depth features; This represents a learnable linear projection bias.

4. The method for non-contact visual settlement and displacement monitoring and operational status assessment of steep slopes according to claim 3, characterized in that, S44 specifically includes: The matching score matrix is ​​calculated and represented as follows: In the formula, Represents the matching score matrix; The dustbin extension is used to explicitly assign unmatched keypoints to each set in the matching score matrix, and the extended score matrix is ​​obtained by adding new rows and columns. It is represented as: In the formula, Represents the first image in image A. i One key point matches the score in the dustbin column; In image B, the first... j One key point matches the score in the dustbin column; This represents the default fractional constant for dustbin; Applying Sinkhorn regularization to the extended fractional matrix Perform regularization iterations to obtain the matching probability matrix. .

5. The method for non-contact visual settlement and displacement monitoring and operational status assessment of steep slopes according to claim 1, characterized in that, In S4, the loss function of the feature matching network model is expressed as: In the formula, This represents the loss of the feature matching network model; This indicates that after Sinkhorn regularization, the th i The key point and the first j The matching probability of each key point; Indicates the first i The probability of a key point matching dustbin; Indicates the first j The probability of a key point matching dustbin; , , Let A, B, C, D, and E represent the set of true matching pairs provided during training, the set of keypoint indices for all points that do not have true corresponding points in image A, and the set of keypoint indices for all points that do not have true corresponding points in image B, respectively.

6. The method for non-contact visual settlement and displacement monitoring and operational status assessment of steep slopes according to claim 1, characterized in that, S6 specifically includes: Based on the keypoint pairs extracted from two adjacent high and steep slope images, the pixel displacement of the keypoints in the camera canvas coordinates is obtained, and the pixel displacement is converted into the actual displacement. Based on the actual displacement, the camera coordinates are converted to obtain the settlement displacement data of the high and steep slope in a unified coordinate system.

Images