Unmanned aerial vehicle inspection method and system applied to foundation pit accumulated water monitoring

Abstract

The invention discloses an unmanned aerial vehicle inspection method and system applied to foundation pit accumulated water monitoring, and the method comprises the steps: determining an optimal route node sequence, and generating an unmanned aerial vehicle control instruction set; acquiring internal and external parameters of a multi-source sensor, and performing time alignment compensation on image frames and sensor data; acquiring a ponding area image coordinate set and attribute data; a three-dimensional projection point set of the ponding area is obtained, and the real coverage area of the ponding area is calculated; generating a three-dimensional overview map of the foundation pit by using the multi-view image of the foundation pit area; determining a three-dimensional coordinate point set and an area estimation result of the ponding area; and projecting the three-dimensional coordinate point set of the ponding area to the three-dimensional overview map of the foundation pit to generate a hidden danger monitoring report. By integrating the unmanned aerial vehicle, the sensor data and the advanced image processing technology, the accuracy, stability and visualization effect of foundation pit accumulated water monitoring are greatly improved, more reliable and comprehensive technical support is provided for safety management of foundation pit engineering, and the safety risk in the construction process is remarkably reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle intelligent inspection, and particularly relates to an unmanned aerial vehicle inspection method and system applied to pit water monitoring. BACKGROUND

[0002] As an important part of underground structure construction, pit engineering has complex construction environment and long construction period, and is easily affected by factors such as climate conditions, hydrological changes and surrounding environment disturbances, especially in rainy seasons or areas with abundant groundwater. The pit water problem is particularly prominent. The water not only may cover the construction risk points and weaken the slope stability, but also may cause abnormal stress of the supporting structure, and even lead to slope instability, pit collapse and other major safety accidents. Therefore, real-time monitoring and accurate grasping of the pit area water condition become an important means to ensure the safe operation and scientific management of pit engineering. At present, the water monitoring methods widely used in engineering mainly include manual inspection, ground sensor layout and fixed camera observation. Such methods have obvious limitations. First, manual inspection depends on the experience of operating personnel, is limited by the inspection frequency, weather conditions and site accessibility, and is difficult to achieve comprehensive and continuous monitoring. Second, the liquid level sensor or water immersion alarm laid on the ground can only detect the set point, cannot cover the entire pit area, and the equipment is easily affected by silt accumulation or electrical failure, resulting in data loss. Although the fixed camera can provide continuous image information, its viewing angle is limited and cannot meet the identification needs of a large area, especially hidden corners. In addition, most of the traditional image recognition methods use binary segmentation or simple threshold processing based on single-frame images, which cannot accurately distinguish between water and shadow, wet mud or reflective areas, and are prone to false positives or false negatives, and it is difficult to extract the real spatial position and area information of the water. In the aspect of unmanned aerial vehicle application, although it has the advantages of flexible deployment and wide coverage, the existing inspection strategies are mostly based on rule-based flight routes, and lack of precise control of pit structure boundaries, flight parameters and image overlap, resulting in poor continuity and incomplete coverage of aerial images, affecting the subsequent image processing effect. At the same time, the existing methods often fail to effectively integrate unmanned aerial vehicle attitude data, laser ranging information and image content, making it difficult to accurately project two-dimensional images into three-dimensional space, thereby limiting the physical quantification and spatial positioning of the water area. In summary, the existing technology still has significant deficiencies in the comprehensive perception, automatic identification, spatial positioning and information integration of pit water, and needs to be systematically improved and enhanced in terms of data acquisition accuracy, target detection stability and result visualization presentation capability. SUMMARY

[0003] The present application provides an unmanned aerial vehicle inspection method and system applied to pit water monitoring to solve the above problems of the existing technology.

[0004] The embodiment of the first aspect of the application provides an unmanned aerial vehicle inspection method applied to foundation pit water accumulation monitoring, mainly comprising: According to the spatial boundary coordinates of the foundation pit area, the flight operation height, the camera field of view angle and the image resolution parameters, an initial route set is constructed, an optimal route node sequence is determined, and a set of unmanned aerial vehicle control instructions is generated; According to the camera calibration board image, the radar ranging residual error and the IMU motion simulation data, multi-source sensor internal and external parameters are obtained, and time alignment compensation of image frames and sensor data is performed; Based on historical foundation pit area image frames and corresponding timestamp information, the segmentation mask of the water accumulation area in the foundation pit area image, the corresponding boundary box information and the confidence score are extracted, and the water accumulation area image coordinate set and attribute data are obtained; According to the pixel mask coordinates of the foundation pit area image and the camera internal parameters, the rotation matrix constructed based on the Euler angle and the unmanned aerial vehicle track position vector are combined to perform world coordinate system conversion, and the real coverage area of the water accumulation area is calculated; Using multi-view images of the foundation pit area, SIFT or ORB local feature points are extracted, cross-view corresponding relationship is constructed, and camera attitude matrix, track position vector and dense point cloud surface fitting results are combined to generate a three-dimensional overview map of the foundation pit; According to the corresponding three-dimensional coordinate point set of the water accumulation area in each image frame, the area coincidence degree of the cross-temporal multi-view segmentation point set is combined to determine the three-dimensional coordinate point set and the area estimation result of the water accumulation area; The three-dimensional coordinate point set of the water accumulation area is projected onto the three-dimensional overview map of the foundation pit, and the center position of the water accumulation area boundary box and the image frame index are combined to generate a hidden danger monitoring report.

[0005] Further, the method comprises the following steps: According to the spatial boundary coordinates of the foundation pit area, the flight operation height, the camera field of view angle and the image resolution parameters, the image ground projection size and the flight line spacing are calculated; by setting the lateral and longitudinal overlap rate threshold, a heading and lateral coverage matrix is constructed; a regular grid scanning path is generated by using a rotating minimum circumscribed rectangle constraint, and a spiral progressive path starting from the center of the foundation pit is constructed based on a polar coordinate generation function; by taking the regular grid scanning path and the spiral progressive path as an initial flight line set, and constructing a multi-objective optimization model with flight path coverage rate and path overlap degree as objective functions, the genetic algorithm is used to iteratively optimize the flight line node sequence and flight attitude angle parameters, and the optimal flight line node sequence meeting the coverage and energy balance requirements is obtained by combining the distance constraint between flight line nodes and the minimum turning radius limit; according to the optimal flight line node sequence, the latitude, longitude, height and shooting trigger time information of each node are obtained, and the unmanned aerial vehicle control instruction set and the unmanned aerial vehicle flight monitoring route are generated; according to the unmanned aerial vehicle flight monitoring route, the camera and the radar are carried on the unmanned aerial vehicle, and the multi-view images of the foundation pit area are shot in real time and stored in the foundation pit area monitoring database.

[0006] Further, the multi-source sensor internal and external parameters are obtained according to the camera calibration board image, the radar ranging residual error and the IMU motion simulation data, and the time alignment and compensation of the image frame and the sensor data are performed, including: By shooting a sequence of calibration board images, the camera internal parameter matrix is calculated by using the Zhang Zhengyou camera calibration method to obtain the focal length parameter, the principal point coordinates and the distortion coefficient; the radar ranging calibration is performed according to the known distance reflection plate, the residual data of the actual distance and the radar ranging output are recorded, the least square regression method is used to establish a ranging error correction model to determine the radar ranging correction coefficient; according to the three-axis angular velocity and linear acceleration data output by the inertial measurement unit, a six-degree-of-freedom motion simulation test is implemented, and the extended Kalman filter algorithm is used to estimate the attitude solution error; by adjusting the off-line parameters of the bias term and the noise covariance matrix, the IMU attitude drift correction model parameters are determined, and the calibrated camera, radar and IMU internal and external parameter data are obtained, the IMU internal parameter data includes zero offset, scale factor and noise characteristics, and the IMU external parameter data includes the pose transformation matrix relative to the camera and the radar; according to the calibrated camera, radar and IMU internal and external parameter data, the unmanned aerial vehicle autonomous flight task is executed, and the spatial position, attitude parameter and shooting trigger timestamp of the unmanned aerial vehicle are obtained, the spatial position includes the flight path position and the radar height information, the attitude parameter includes the attitude angle and the flight speed, and the attitude angle includes the roll angle, the pitch angle and the yaw angle; by comparing the shooting trigger timestamp with the timestamp recorded by other sensors, the multi-source data time compensation is performed by using the linear interpolation alignment method to obtain the spatial position and attitude parameter set at the corresponding image frame time.

[0007] Further, based on the historical foundation pit area image frames and corresponding timestamp information, the segmentation mask of the water accumulation area in the foundation pit area image, the corresponding boundary box information and the confidence score are extracted, and the water accumulation area image coordinate set and attribute data are obtained, including: Through the foundation pit area monitoring database, the historical foundation pit area image frames and corresponding timestamp information are obtained, the image enhancement processing method of random cropping, rotation, multi-scale scaling and brightness contrast adjustment is adopted, the extended image sample set containing multiple perspectives, different lighting conditions and different scale changes is generated; the water accumulation area of the historical foundation pit area image and the extended image sample set is drawn with a rectangular boundary box in a box selection labeling manner, the yolov11 target detection algorithm is used for model training, the segmentation mask of the water accumulation area in the foundation pit area image, the corresponding boundary box information and the confidence score are extracted; the extracted water accumulation area image is decoded to obtain the pixel mask coordinate point set of the water accumulation area in the image coordinate system; the area pixel number and the boundary polygon vertex coordinates of each water accumulation area are counted, the pixel-level boundary contour of the water accumulation area is calculated, the image frame index, the mask coordinates, the confidence and the timestamp information are stored to the foundation pit area monitoring database, and the image coordinate set and attribute information of the water accumulation area are obtained; through the pre-set confidence threshold and mask integrity screening condition, the pseudo-detection area and the boundary damaged target are removed, the target mask set meeting the spatial continuity requirement is retained, and the pixel mask coordinate point set and the image frame index information are recorded.

[0008] Further, according to the pixel mask coordinates of the foundation pit area image and the camera intrinsic parameters, the rotation matrix constructed by the Euler angle and the unmanned aerial vehicle track position vector are combined to perform world coordinate system conversion and calculate the real coverage area of the water accumulation area, including: According to the pixel mask coordinates of the foundation pit area image and the camera intrinsic parameters, the water accumulation area boundary points in the image coordinate system are mapped to the space rays in the camera coordinate system by using the pinhole imaging model, and the formula is wherein, is the space ray in the camera coordinate system, is the scale factor, is the camera intrinsic matrix, , is the focal length, , is the principal point coordinate, is the image pixel point coordinate; according to the roll angle, the pitch angle and the yaw angle recorded in the flight process of the unmanned aerial vehicle, the Euler angle split shaft rotation method is adopted to calculate the rotation matrix around the X axis, the Y axis and the Z axis in turn, and a 3×3 space rotation matrix is formed by matrix multiplication combination; based on the rotation matrix and the track position vector of the unmanned aerial vehicle, the camera coordinate system ray is converted to the world coordinate system, and the formula is, wherein, is the coordinate in the world coordinate system, is the attitude matrix, is the track position vector; the two-dimensional pixel mask is projected to three-dimensional coordinates, and the real coverage area of the waterlogging area is calculated by triangulation or gridding method wherein, is the projected area grid of a single waterlogging segmentation mask, and N is the number of grid points of the mask.

[0009] Further, the multi-view image of the foundation pit area is used to extract SIFT or ORB local feature points, to construct a cross-view correspondence relationship, and to generate a three-dimensional overview map of the foundation pit by combining the camera attitude matrix, the track position vector, and the dense point cloud surface fitting result, including: The multi-view image of the foundation pit area is used to extract SIFT or ORB local feature points, to establish a feature correspondence relationship between multi-view images; the structure self-motion method is used to solve the attitude matrix and the track position vector of the image frame in the world coordinate system, and to restore the sparse three-dimensional point cloud , the formula is wherein, represents a pinhole projection function, which maps three-dimensional points to an image plane, is the number of feature points of the image, is the number of image frames; according to the laser radar height ranging value recorded simultaneously during flight, the Z-axis coordinates of the three-dimensional sparse point cloud are aligned with the radar ground height reference, and the formula is used to correct the Z-axis of the three-dimensional sparse point cloud, wherein, is the average height of the sparse point cloud; according to the camera pose parameters solved by the structure self-motion method and the three-dimensional sparse point cloud result, a multi-view stereo matching algorithm is used to estimate the disparity and depth of each pixel in the image, to construct a high-density three-dimensional dense point cloud, and the pose parameters include spatial position and attitude parameters; a Poisson reconstruction or Delaunay triangulation algorithm is used to surface fit and topological model the dense point cloud, to generate a three-dimensional overview map of the foundation pit.

[0010] Further, the three-dimensional coordinate point set and area estimation result of the waterlogging area are determined according to the corresponding three-dimensional coordinate point set of the waterlogging area in each image frame, combined with the area coincidence degree of the cross-temporal multi-view segmentation point set, including: According to all the pixel points in the real coverage area of the waterlogging area of each acquisition frame, the corresponding three-dimensional coordinate points are converted through the radar position and pose information at the corresponding moment; the segmentation coordinate point set of multi-view and multi-time in time sequence is expressed as , the area coincidence degree of different detection results on the reference plane is calculated by the formula , and it is not difficult to remove the detection results with a coincidence degree less than the preset coincidence degree threshold a low confidence detection result less than a preset coincidence degree threshold, wherein, a detection result with a coincidence ratio on a projection plane, , respectively, three-dimensional coordinate point sets of two detection results; according to the remaining high-confidence detection result, the three-dimensional coordinate point set of the waterlogging area is determined using the formula and an area estimation result , wherein, is a weight, which is obtained by comprehensively calculating the segmentation confidence, the area coincidence degree or the multi-view coverage degree, is the area of the i-th detection area.

[0011] Further, the hidden danger monitoring report is generated by projecting the three-dimensional coordinate point set of the waterlogging area to the three-dimensional overview map of the foundation pit, and combining the center position of the bounding box of the waterlogging area and the image frame index. According to the three-dimensional coordinate point set of the waterlogging area, the grid structure, the face normal vector and the positional relationship of the three-dimensional overview map of the foundation pit, the nearest surface area corresponding to each waterlogging point is determined, and the projection position of the waterlogging area in the three-dimensional overview map of the foundation pit is determined using the vertical projection, the normal projection or the shortest distance matching strategy. The three-dimensional coordinate point set of the waterlogging area is projected into the three-dimensional overview map of the foundation pit; according to the projection position of the waterlogging area in the three-dimensional overview map of the foundation pit, the number of waterlogging targets in the current task period is determined, and each target is assigned a unique number identifier; based on the center position of the bounding box of the three-dimensional coordinate point set of the waterlogging area, the three-dimensional position coordinates of the waterlogging area in the world coordinate system are determined, and the original image frame index corresponding to the waterlogging area is obtained. The image cropping area containing the segmentation mask of the waterlogging area is extracted as image evidence; according to the number, the number, the three-dimensional position coordinates, the area and the corresponding image evidence of the waterlogging area, a hidden danger monitoring report of the foundation pit area is generated and stored in the foundation pit area monitoring database.

[0012] The second aspect embodiment of the present application provides an unmanned aerial vehicle inspection system applied to foundation pit waterlogging monitoring, mainly comprising: An optimal flight monitoring route generation module is used to construct an initial route set according to the spatial boundary coordinates of the foundation pit area, the flight operation height, the camera field of view angle and the image resolution parameters, determine an optimal route node sequence, and generate an unmanned aerial vehicle control instruction set. A multi-source sensor calibration and time sequence alignment module is used to obtain multi-source sensor internal and external parameters according to the camera calibration board image, the radar ranging residual error and the IMU motion simulation data, and perform time alignment compensation of image frames and sensor data. ​A water accumulation area identification module is configured to extract a segmentation mask of a water accumulation area, corresponding bounding box information and a confidence score in a foundation pit area image based on historical foundation pit area image frames and corresponding timestamp information, to obtain a water accumulation area image coordinate set and attribute data; A three-dimensional projection area estimation module is configured to perform world coordinate system conversion and calculate a real coverage area of a water accumulation area according to pixel mask coordinates of a foundation pit area image and camera intrinsic parameters, in combination with a rotation matrix constructed based on Euler angles and a UAV track position vector. A foundation pit three-dimensional overview map construction module is configured to extract SIFT or ORB local feature points by using multi-view images of a foundation pit area, to construct a cross-view correspondence relationship, and to generate a foundation pit three-dimensional overview map in combination with a camera pose matrix, a track position vector and a dense point cloud surface fitting result. A multi-view water accumulation point set fusion module is configured to determine a three-dimensional coordinate point set and an area estimation result of a water accumulation area according to corresponding three-dimensional coordinate point sets of the water accumulation area in each image frame, in combination with an area coincidence degree of cross-temporal multi-view segmentation point sets. A hidden danger monitoring report generation module is configured to project the three-dimensional coordinate point set of the water accumulation area to the foundation pit three-dimensional overview map, and to generate a hidden danger monitoring report in combination with a bounding box center position of the water accumulation area and an image frame index.

[0013] The technical scheme provided by the embodiments of the present application can have the following beneficial effects: The application provides a kind of unmanned aerial vehicle inspection method and system applied to foundation pit waterlogging monitoring.The application can automatically generate optimal route sequence by combining the parameters such as spatial boundary of foundation pit area, flight operation height, camera field of view angle and resolution, ensure to cover the entire foundation pit area, overcome the problem of limited coverage and more blind area of traditional monitoring method.Secondly, based on time alignment and compensation of multi-source sensor data, high-precision image and sensor data fusion is realized, which greatly improves the accuracy and stability of monitoring data.Through processing of historical image data and accurate segmentation and attribute extraction of waterlogging area, the application can identify waterlogging area in real time and generate corresponding spatial coordinate data, to ensure accurate positioning of waterlogging information.Combined with three-dimensional spatial information and multi-view images of foundation pit area, the application can realize three-dimensional projection of waterlogging area and physical quantization of coverage area, improve the spatial performance and visualization ability of monitoring results.On this basis, by removing low-confidence detection results through coincidence threshold, the accuracy and reliability of waterlogging area detection are improved.The application realizes comprehensive perception, accurate identification and real-time monitoring of foundation pit waterlogging by projecting waterlogging area and foundation pit three-dimensional overview map in space to generate hazard monitoring report, effectively ensures the safety and stability of foundation pit engineering.The application greatly improves the accuracy, stability and visualization effect of foundation pit waterlogging monitoring by integrating unmanned aerial vehicle, sensor data and advanced image processing technology, provides more reliable and comprehensive technical support for safety management of foundation pit engineering, and significantly reduces the safety risk in construction process. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 A flowchart of the unmanned aerial vehicle inspection method applied to foundation pit waterlogging monitoring of the application; Figure 2 A schematic diagram of the unmanned aerial vehicle inspection method applied to foundation pit waterlogging monitoring of the application; Figure 3 A schematic diagram of the unmanned aerial vehicle inspection system applied to foundation pit waterlogging monitoring of the application. DETAILED DESCRIPTION

[0015] To make the purpose, technical scheme and advantages of the application clearer, the application will be described in detail below in combination with drawings and specific embodiments, obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments.

[0016] As shown in Figure 1 The overall flowchart of the unmanned aerial vehicle inspection method applied to foundation pit waterlogging monitoring of the application, which shows the execution order and logical relationship of core processing steps.Further, as Figure 2As shown, it is a structural schematic diagram of the system of the present application, which details the composition of each functional module and the data flow and calculation relationship between them, including image instance segmentation processing, three-dimensional space mapping, time sequence fusion and report generation, and other key links, thus fully embodying the technical implementation path of the present application.

[0017] In the embodiment of the present application, before flight operation, the flight operator needs to understand the whole area of the monitored foundation pit and set the operation boundary range. According to the two-dimensional coordinate point set of the foundation pit boundary, the preset flight height, the camera field of view angle and the image resolution parameters, the projection size of a single frame image on the ground is calculated. By setting the lateral and longitudinal overlap rate threshold, the flight line spacing is calculated and the coverage matrix is constructed, forming the model of the image overlap relationship in the direction of flight and the lateral direction. Based on the foundation pit boundary contour, the regular grid scanning path is generated by using the rotating minimum circumscribed rectangle method, and the spiral progressive path with the geometric center of the foundation pit as the starting point is constructed by using the polar coordinate generation function. The regular grid scanning path and the spiral progressive path are used as the initial flight line set, a multi-objective optimization model is constructed with the maximum flight path coverage rate and the minimum path overlap rate as the objective function, the genetic algorithm is used to iteratively solve the flight line node sequence and the flight attitude angle parameters, and the optimal flight monitoring flight line node sequence meeting the coverage and energy balance requirements is obtained by combining the distance constraint between nodes, the minimum turning radius and the path continuity constraint. According to the optimized flight line sequence, the unmanned aerial vehicle flight control instruction set containing the latitude, longitude, height, heading angle, flight speed and shooting trigger time is generated for each node, and the flight monitoring flight line is automatically constructed. During the flight task execution process, the task instructions are issued to the unmanned aerial vehicle, the unmanned aerial vehicle is controlled to automatically fly along the planned path and simultaneously perform image acquisition. The unmanned aerial vehicle platform needs to carry a high-definition image acquisition module, a laser radar and an inertial measurement unit. The camera acquires the camera intrinsic matrix including focal length , principal point coordinates and radial distortion coefficient by shooting a standard calibration board image sequence, which is used for subsequent projection conversion of image pixel coordinates to space rays. The laser radar is used to provide height ranging , and the IMU records the attitude angle and acceleration / angle velocity information, auxiliary flight solution, by setting multiple groups of different known distance reflector plates, comparing the residual error of ranging output and actual distance, using least square method to construct one-dimensional fitting regression model, determining the ranging error correction coefficient, and applying it to the height correction during flight. The IMU records the output of three-axis angular velocity and linear acceleration through six-degree-of-freedom motion simulation experiment, estimates the attitude solution error by using extended Kalman filtering method, optimizes the noise covariance matrix and zero offset model parameters offline, and obtains the IMU attitude drift correction model. Finally, the camera, radar and IMU are calibrated, the IMU internal parameters include zero offset, scale factor and measurement noise characteristics, and the external parameters include the relative pose transformation matrix of camera and radar. When the unmanned aerial vehicle executes the task according to the flight route, the image , track position , attitude matrix and radar ranging value are collected in real time, the timestamp generated by the image acquisition trigger signal is compared with the timestamp in the sensor data stream, the linear interpolation method is used to complete the time sequence alignment and synchronization compensation of multi-source data of image, IMU and radar, the spatial pose data set at image frame level is constructed, and the data set is uploaded to the unmanned aerial vehicle inspection platform through the communication link. The platform is the core node of the whole platform, integrates communication access, data scheduling, processing control and result uploading functions, is responsible for coordinating the operation of each module, ensures the real-time receiving, processing and storage of flight data, image segmentation results and three-dimensional projection and area estimation information, and realizes the unified management and visual presentation of the whole process of foundation pit water monitoring.

[0018] The instance segmentation framework accessed in the platform is used for automatic identification processing of the foundation pit water accumulation image. The framework takes YOLOv11 model as the core detection algorithm, and can realize target segmentation, boundary box extraction and confidence score output of the water accumulation area in the unmanned aerial vehicle collected image. In order to improve the generalization ability and positioning accuracy of the model, the training data set is derived from the historical image frames stored in the foundation pit area monitoring database and the corresponding collection timestamp information. The historical image frames are labeled and expanded to form a sample set for model training. Before training, first, perform enhancement processing operation on the historical image frames, specifically including applying random cropping, affine rotation, multi-scale scaling and brightness contrast adjustment strategy to the original image to generate image expansion samples containing multi-view, lighting conditions and size changes. The artificial auxiliary frame selection labeling method is used to draw the rectangular boundary box of the water accumulation area in the original image and the enhanced image, and the annotation file is generated according to the YOLO input format. In the training process, the model backbone network adopts CSPDarknet architecture to extract multi-scale spatial features, introduces FPN+PAN pyramid structure in the Neck part for feature fusion, and the Head module adopts a multi-head prediction mechanism to output boundary boxes, masks and classification confidence at different down-sampling rates. The model training adopts a combined loss function strategy, which jointly uses GIoU Loss, Focal Loss, Dice Loss and BCE Loss to optimize the boundary, class and mask accuracy. The optimizer adopts Adam structure, cooperates with learning rate preheating and cosine annealing scheduling mechanism to improve the model convergence rate and stability. After training, the model is deployed in the platform inference engine to perform prediction on the unmanned aerial vehicle collected image. The platform performs post-processing operation on the segmentation mask output by each image, extracts the pixel coordinate point set of the mask boundary polygon, calculates the pixel area, boundary contour and confidence score of each mask, and writes the image frame index, mask coordinate point set, boundary box position information and timestamp field into the foundation pit area monitoring database. According to the dynamic water level and soil type of each mask area, the pre-set confidence threshold of the area is adaptively adjusted. Through the pre-set confidence threshold and the mask connectivity judgment condition, the abnormal detection results with low confidence prediction and boundary structure damage are removed, and only the target mask set with continuous boundary and area greater than the minimum area threshold is retained. The actual deployment results show that under the condition of 1280x1280 resolution image input, the average prediction time of a single image can be controlled within hundreds of milliseconds, and the GPU accelerated deployment environment can realize multi-channel image parallel inference. In multiple foundation pit scene tests, the target segmentation accuracy reaches more than 90%, and the platform can stably output water accumulation area mask image, boundary coordinates, confidence and image frame number and other key information, meeting the comprehensive requirements of accuracy, speed and traceability in foundation pit inspection tasks.

[0019] The specific steps for adaptively adjusting the preset confidence threshold for each masked area of ​​the foundation pit, based on the dynamic water level and soil type, are as follows: The images acquired by the UAV are converted from RGB to HSV color space. The mean, variance, and spatial continuity indices of chromaticity are statistically analyzed to obtain the color feature vectors of each area of ​​the foundation pit. The boundaries and color gradient fields of each foundation pit area are extracted through sub-pixel edge detection. The gray-level co-occurrence matrix, local binary mode, and histogram of directional gradients are calculated to obtain the texture feature vectors of each foundation pit area, which are then stored in the foundation pit area monitoring database. Through the foundation pit area monitoring database, historical color feature vectors and texture feature vectors of the foundation pit areas are obtained, and the soil type of each area is labeled. A random forest algorithm is used for model training to construct a soil type recognition model for foundation pit areas, identifying the soil type of each masked area within the foundation pit, thereby determining the soil type of each masked area. The water accumulation area within the foundation pit changes dynamically with the groundwater level, and the soil type has a significant impact on the water's reflectivity. Real-time water level changes within the foundation pit are obtained using radar ranging data. The radar ranging system measures the distance between the water surface and the radar by emitting electromagnetic waves and receiving their reflected signals, thus providing accurate water level data. The water level change Zt over time t can be expressed as... ,in, This represents the initial water level of the foundation pit. The change in water level over time t represents the dynamic fluctuation of the water depth in the foundation pit. Meanwhile, soil type significantly affects the reflectivity of the masking area; different soil types, such as clay, sandy soil, and loam, result in varying degrees of light reflection. A soil influence coefficient is defined based on soil type s. To correct the reflection characteristics of the masked region, an adaptive confidence threshold correction formula is used. Adaptively adjust the preset confidence threshold for the mask region. This makes the detection of the masked area more accurate, among which, This is the initial detection threshold. This is a correction factor for the detection threshold based on soil type. Based on the adjusted preset confidence threshold, waterlogged areas are identified. If the confidence level of the currently extracted waterlogged area is greater than the adjusted preset confidence threshold, the area is classified as a waterlogged area; otherwise, it is considered a background area.

[0020] Next, the pixel mask data of the water accumulation area needs to be processed by the spatial projection and area estimation module to achieve the mapping transformation from image coordinates to real three-dimensional space. The UAV inspection platform uses the boundary point coordinates of the water accumulation target pixel mask in each frame of the image and the calibrated camera intrinsic parameters, including focal length, to perform the mapping. Principal point coordinates and radial distortion coefficient Using a pinhole imaging model to capture image pixels Mapped to a spatial ray in the camera coordinate system The mapping is expressed as ,in, The scale factor is determined by depth information or radar ranging. Subsequently, based on the timestamp of the image acquisition time, the platform aligns the IMU output with the flight track record using linear interpolation to obtain the UAV flight attitude data corresponding to the current frame image, including three-axis attitude angles. Using angular velocity data, rotation matrices around the X, Y, and Z axes are constructed sequentially using Euler angle-separated rotation. These matrices are then combined through matrix multiplication to generate the rotation matrix from the camera to the world coordinate system. Combined with real-time recorded track position vectors The spatial transformation process is completed, converting the spatial ray from the camera coordinate system to the world coordinate system. The mapping relationship is as follows: ,in, The coordinates are in the world coordinate system. After the projection transformation, the platform extends the spatial rays generated by the projection of each mask boundary point to the plane where the target is located. Based on the radar altitude ranging information or the terrain surface Z-value provided by the 3D overview model of the foundation pit, the coordinates of the intersection points of the rays and the ground are extracted to form a 3D point set of the water accumulation area. The platform uses a triangular meshing method to construct the contour according to the point set order, converting the mask boundary into a closed 3D projected boundary, forming an area estimation mesh structure. The projected area corresponding to a single mask area is calculated using the formula: ,in, The projected area grid of a single water accumulation mask is divided, where N is the number of grid points of the mask. The area is obtained by accumulating grid points one by one. Finally, the platform obtains the coverage area data of each water accumulation target at the real scale.

[0021] In this embodiment, to present the monitoring results in a more concrete and unified manner, the platform generates a 3D overview of the foundation pit based on multi-view aerial images from UAVs, flight path positions, attitude parameters, and radar altitude ranging information. Firstly, multi-view images are used... Extract local feature points from SIFT or ORB. ,in This represents the j-th feature point in the i-th image, and cross-view correspondence is determined through feature matching. The camera pose was calculated using the Structure Self-Motion (SfM) method. And restore sparse 3D point clouds The formula is ,in This represents the pinhole projection function, which maps 3D points onto the image plane. For the first The number of feature points in the image. To eliminate scale uncertainty in SfM, radar altitude ranging is introduced. The Z-axis of the point cloud is corrected, and the formula is , wherein is the average height of the sparse point cloud. Subsequently, the sparse point cloud is expanded into a dense point cloud by using a dense matching algorithm such as multi-view stereo matching. The dense point cloud is expanded into a dense point cloud by using a surface reconstruction algorithm such as Poisson reconstruction or Delaunay triangulation, and a three-dimensional overview of the foundation pit with terrain height and structural continuity is generated , and the formula is The three-dimensional overview of the foundation pit truly reflects the geometric structure and spatial form of the foundation pit, and can be used as a unified reference plane for subsequent three-dimensional projection, area calculation and time sequence result fusion of the water accumulation area.

[0022] In the embodiments of the present application, the segmentation prediction is performed on images that are discrete in time sequence, so the results contain repeated multi-views or confused misjudgments in some views. A related time sequence multi-view fusion framework is arranged on the platform. The predicted area of each acquisition frame contains all the pixel points which are converted into corresponding three-dimensional coordinate points by using the radar tracking and pose information at the corresponding moment. At this time, the segmentation coordinate point set in the time sequence multi-view and multi-moment can be expressed as For the overlapping area between different detection results, the area coincidence degree in the reference plane is calculated, wherein represents the coincidence ratio of the detection result and in the projection plane. This process is expanded and paired in the time sequence, and multiple exist. It can be determined as a low-confidence accidental detection and eliminated. For the remaining high-confidence detection results, the platform obtains the final accurate target position and area estimation based on a weighted fusion method, wherein is the weight, which can be calculated comprehensively by the segmentation confidence, the area coincidence degree or the multi-view coverage degree, is the area of the m-th detection area. Finally, the fused result is accurately projected onto the three-dimensional overview of the foundation pit by using the surface projection function , so as to realize the unified display and analysis of multi-view and multi-time sequence results. Through this process, the present application can maintain the consistency and robustness of the detection results in a complex environment, and provide accurate three-dimensional positioning and area evaluation for water accumulation monitoring.

[0023] In the implementation of the above method, the platform realizes the alignment mapping and unique identification of the monitoring target in the three-dimensional model based on the constructed three-dimensional overview of the foundation pit and the three-dimensional coordinate point set of the water accumulation area through the spatial geometric projection module. Specifically, according to the three-dimensional point set coordinates of the water accumulation area and the established grid structure, face normal vector and spatial position relationship in the three-dimensional model of the foundation pit, the platform successively judges the most adjacent surface area corresponding to each water accumulation point in the three-dimensional terrain surface, and preferentially adopts the normal projection strategy to realize the projection mapping of the three-dimensional point set to the grid surface. When the normal direction does not converge or the face inclination angle exceeds the set threshold, the system automatically switches to the vertical projection or the shortest distance matching strategy to complete the position alignment of the point set in the three-dimensional overview. After completing the point set mapping, the platform performs spatial clustering on the water accumulation targets according to the face attribution result, and counts the effective cluster number of all water accumulation targets in the current monitoring period to generate a corresponding number of hidden danger target entities. The platform automatically assigns a unique number to each cluster area and records the three-dimensional projection area range corresponding to the number to form a number-coordinate mapping relationship. In the coordinate labeling process, the platform extracts the geometric center position of each cluster area based on the three-dimensional bounding box structure of the cluster area as the spatial positioning information of the water accumulation area in the world coordinate system. At the same time, the platform extracts the corresponding image frame from the image database in combination with the original image frame index associated with the cluster number, obtains the local image area containing the target through boundary index cropping, and superimposes the segmentation mask as image evidence to improve the verifiability and image reproducibility of the monitoring result. Finally, the platform organizes the number identification, three-dimensional position coordinates, coverage area data and image evidence of each water accumulation target into a standardized data structure, automatically generates a structured hidden danger monitoring report, and completes uploading to the foundation pit area monitoring database to realize the functions of unified archiving, historical backtracking and horizontal comparison of monitoring data.

[0024] As Figure 3 The unmanned aerial vehicle inspection system for foundation pit water accumulation monitoring according to the embodiment can specifically include: An optimal flight monitoring route generation module is configured to construct an initial route set according to the spatial boundary coordinates of the foundation pit area, the flight operation height, the camera field of view angle and the image resolution parameters, determine an optimal flight monitoring route node sequence in combination with the route node sequence and the flight attitude angle parameters, and generate a set of unmanned aerial vehicle control instructions. A multi-source sensor calibration and time alignment module is configured to obtain multi-source sensor internal and external parameters according to the camera calibration board image, the radar ranging residual error and the IMU motion simulation data, and perform time alignment compensation on the image frames and sensor data. The water accumulation area identification module is configured to perform image enhancement and boundary box labeling based on historical foundation pit area images and corresponding timestamp information in a foundation pit area monitoring database, extract a segmentation mask of a water accumulation area in the foundation pit area image, corresponding boundary box information and a confidence score, and obtain a water accumulation area image coordinate set and attribute data. The three-dimensional projection area estimation module is configured to perform world coordinate system conversion based on pixel mask coordinates of the foundation pit area image and camera intrinsic parameters, in combination with a rotation matrix constructed based on Euler angles and a UAV flight path position vector, to generate a three-dimensional projection point set of the water accumulation area and calculate a real coverage area of the water accumulation area. The foundation pit three-dimensional overview map construction module is configured to extract SIFT or ORB local feature points based on multi-view images of the foundation pit area, construct a cross-view correspondence relationship, and generate a foundation pit three-dimensional overview map in combination with a camera pose matrix, a flight path position vector and a dense point cloud surface fitting result. The multi-view water accumulation point set fusion module is configured to perform overlap threshold elimination on low-confidence detection results based on an area overlap degree of cross-temporal multi-view segmentation point sets, and determine a three-dimensional coordinate point set and an area estimation result of the water accumulation area. The hidden danger monitoring report generation module is configured to project the three-dimensional coordinate point set of the water accumulation area to the foundation pit three-dimensional overview map based on a spatial correspondence relationship between the three-dimensional coordinate point set of the water accumulation area and a normal vector of a mesh patch of the foundation pit three-dimensional overview map, and generate a hidden danger monitoring report in combination with a boundary box center position of the water accumulation area and an image frame index.

[0025] The above description is merely preferred embodiments of the present application and a description of the principles of the technology used. It should be understood by those skilled in the art that the scope of the application disclosed in the present application is not limited to the technical solutions formed by the specific combinations of the technical features described above, and also covers other technical solutions formed by any combination of the above technical features or equivalent features without departing from the concept of the present application. For example, the above features can be replaced with technical features disclosed in the present application (but not limited to) having similar functions to form technical solutions.

Claims

1. A drone inspection method for monitoring water accumulation in foundation pits, characterized in that, The method includes: Based on the spatial boundary coordinates of the pit area, flight operation altitude, camera field of view and image resolution parameters, an initial flight path set is constructed, the optimal flight path node sequence is determined, and a UAV control command set is generated. Based on the camera calibration board image, radar ranging residual and IMU motion simulation data, multi-source sensor intrinsic and extrinsic parameters are obtained, and time alignment compensation between image frames and sensor data is performed. Based on historical foundation pit area image frames and corresponding timestamp information, the segmentation mask, corresponding bounding box information and confidence score of the water accumulation area in the foundation pit area image are extracted to obtain the image coordinate set and attribute data of the water accumulation area. Based on the pixel mask coordinates of the image of the foundation pit area and the camera intrinsic parameters, combined with the rotation matrix constructed by Euler angles and the UAV track position vector, the world coordinate system transformation is performed, and the true coverage area of ​​the water accumulation area is calculated. Using multi-view images of the foundation pit area, local feature points of SIFT or ORB are extracted, cross-view correspondence is constructed, and combined with camera attitude matrix, track position vector and dense point cloud surface fitting results, a three-dimensional overview map of the foundation pit is generated. Based on the set of three-dimensional coordinate points corresponding to the water accumulation area in each image frame, and combined with the area overlap of the cross-temporal multi-view segmentation point set, the three-dimensional coordinate point set and area estimation results of the water accumulation area are determined. By projecting the three-dimensional coordinate point set of the water accumulation area onto the three-dimensional overview map of the foundation pit, and combining the center position of the boundary box of the water accumulation area with the image frame index, a hazard monitoring report is generated.

2. The method according to claim 1, wherein, The process involves constructing an initial flight path set based on the spatial boundary coordinates of the pit area, flight operation altitude, camera field of view, and image resolution parameters; determining the optimal flight path node sequence; and generating a UAV control command set, including: Based on the spatial boundary coordinates of the foundation pit area, flight operation altitude, camera field of view, and image resolution parameters, the ground projection size of the image and the spacing between flight paths are calculated. By setting thresholds for lateral and longitudinal overlap rates, a heading and lateral coverage matrix is ​​constructed. A regular raster scanning path is generated using the minimum bounding rectangle constraint of rotation, and a spiral progressive path starting from the center of the foundation pit is constructed based on a polar coordinate generation function. By using the regular raster scanning path and the spiral progressive path as the initial flight path set, a multi-objective optimization model is constructed with track coverage and path overlap as objective functions. A genetic algorithm is used to iteratively optimize the sequence of flight path nodes and flight attitude angle parameters. Combined with distance constraints between flight path nodes and minimum turning radius limits, the optimal flight monitoring flight path node sequence that meets the requirements of coverage and energy consumption balance is obtained. Based on the optimal flight monitoring flight path node sequence, the latitude, longitude, altitude, and shooting trigger time information of each node are obtained to generate a UAV control command set and a UAV flight monitoring flight path. Based on the UAV flight monitoring flight path, a camera and radar are mounted on the UAV to capture multi-view images of the foundation pit area in real time and store them in the foundation pit area monitoring database.

3. The method according to claim 1, wherein, The process of acquiring multi-source sensor intrinsic and extrinsic parameters based on camera calibration board images, radar ranging residuals, and IMU motion simulation data, and performing time alignment compensation between image frames and sensor data, includes: By capturing a sequence of images of the calibration board, the camera intrinsic parameter matrix was calculated using Zhang Zhengyou's camera calibration method to obtain focal length parameters, principal point coordinates, and distortion coefficients. Radar ranging calibration was performed based on a known distance reflector, recording the residual data between the actual distance and the radar ranging output. A ranging error correction model was established using the least squares regression method to determine the radar ranging correction coefficients. Based on the three-axis angular velocity and linear acceleration data output by the inertial measurement unit (IMU), a six-degree-of-freedom motion simulation test was conducted, and the attitude calculation error was estimated using an extended Kalman filter algorithm. By offline adjustment of the bias term and noise covariance matrix parameters, the parameters of the IMU attitude drift correction model were determined, and the calibrated camera, radar, and IMU parameters were obtained. The system uses internal and external parameter data. IMU intrinsic data includes zero bias, scaling factor, and noise characteristics, while IMU extrinsic data includes pose transformation matrices relative to the camera and radar. Based on the calibrated camera, radar, and IMU intrinsic and extrinsic parameter data, the system performs autonomous flight missions and acquires the UAV's spatial position, attitude parameters, and shooting trigger timestamp. Spatial position includes track position and radar altitude information, while attitude parameters include attitude angles and flight speed. Attitude angles include roll, pitch, and yaw angles. By comparing the shooting trigger timestamp with timestamps recorded by other sensors, linear interpolation alignment is used for multi-source data time compensation to obtain the spatial position and attitude parameter set for the corresponding image frame.

4. The method according to claim 1, wherein, Based on historical foundation pit area image frames and corresponding timestamp information, the segmentation mask, corresponding bounding box information, and confidence score of the water accumulation area in the foundation pit area image are extracted to obtain the image coordinate set and attribute data of the water accumulation area, including: By utilizing a foundation pit area monitoring database, historical foundation pit area image frames and corresponding timestamp information were obtained. Image enhancement processing methods, including random cropping, rotation, multi-scale scaling, and brightness / contrast adjustment, were employed to generate an expanded image sample set containing multiple viewpoints, different lighting conditions, and varying scales. Using bounding box annotation, rectangular bounding boxes were drawn for the water-filled areas in both the historical foundation pit area images and the expanded image sample set. The YOLOv11 object detection algorithm was used for model training to extract the segmentation mask, corresponding bounding box information, and confidence score of the water-filled areas in the foundation pit area images. Based on the extracted water-filled areas... The domain image is decoded to obtain the pixel mask coordinates of the water accumulation area in the image coordinate system. By counting the area pixels and the vertex coordinates of the boundary polygons of each water accumulation area, the pixel-level boundary contour of the water accumulation area is calculated. The image frame index, mask coordinates, confidence level and timestamp information are stored in the pit area monitoring database to obtain the image coordinate set and attribute information of the water accumulation area. By setting a confidence threshold and mask integrity screening conditions, false detection areas and targets with broken boundaries are eliminated, and the target mask set that meets the spatial continuity requirements is retained. The pixel mask coordinates and image frame index information are recorded.

5. The method according to claim 1, wherein, The process involves transforming the world coordinate system based on the pixel mask coordinates of the pit area image and camera intrinsic parameters, combined with a rotation matrix constructed using Euler angles and the UAV's flight path position vector, and calculating the actual coverage area of ​​the waterlogged area. This includes: Based on the pixel mask coordinates of the image of the foundation pit area and the camera intrinsic parameters, the boundary points of the water accumulation area in the image coordinate system are mapped to spatial rays in the camera coordinate system using a pinhole imaging model. The formula is as follows: ,in, A spatial ray in the camera coordinate system. As a scale factor, For the camera intrinsic parameter matrix, , Focal length , Principal point coordinates The coordinates of the image pixels are used. Based on the roll, pitch, and yaw angles recorded during the UAV's flight, rotation matrices around the X, Y, and Z axes are calculated sequentially using Euler angle-based rotation. These matrices are then combined through matrix multiplication to form a 3×3 spatial rotation matrix. The rotation matrix is ​​then used to... and track position vector The formula for transforming the camera coordinate system ray to the world coordinate system is: ,in, Coordinates in the world coordinate system Let be the attitude matrix. The trajectory position vector is used; the two-dimensional pixel mask is projected onto three-dimensional coordinates, and the actual coverage area of ​​the water accumulation area is calculated using triangulation or meshing methods. ,in, The projected area grid of the Mask is divided into segments for a single water accumulation, where N is the number of grid points in the Mask.

6. The method according to claim 1, wherein, The process involves using multi-view images of the foundation pit area to extract local feature points using SIFT or ORB, constructing cross-view correspondences, and combining these with camera attitude matrix, track position vector, and dense point cloud surface fitting results to generate a 3D overview map of the foundation pit, including: Using multi-view images of the foundation pit area, local feature points are extracted using SIFT or ORB methods to establish feature correspondences between the multi-view images. The structure self-motion method is employed to calculate the attitude matrix and track position vector of the image frames in the world coordinate system, and to reconstruct the sparse 3D point cloud. The formula is ,in, This represents the pinhole projection function, which maps 3D points onto the image plane. For the first Number of feature points in an image The number of image frames; based on the lidar altitude ranging values ​​recorded during flight, the Z-axis coordinates of the 3D sparse point cloud are aligned with the radar ground altitude reference, and the formula is used. The Z-axis of the 3D sparse point cloud is corrected, whereby... The average height of the sparse point cloud is given. Based on the camera pose parameters calculated by the structure self-motion method and the 3D sparse point cloud results, a multi-view stereo matching algorithm is used to perform disparity estimation and depth inversion on each pixel in the image to construct a high-density 3D dense point cloud. The pose parameters include spatial position and attitude parameters. The dense point cloud is then surface-fitted and topologically modeled using Poisson reconstruction or Delaunay triangulation algorithms to generate a 3D overview of the foundation pit.

7. The method according to claim 1, wherein, The process of determining the three-dimensional coordinate point set and area estimation result of the water accumulation area based on the set of three-dimensional coordinate points corresponding to the water accumulation area in each image frame, combined with the area overlap of the cross-temporal multi-view segmentation point set, includes: Based on all pixels in the actual coverage area of ​​the water accumulation region in each acquisition frame, and using the radar and pose information at the corresponding time moment, they are converted into corresponding 3D coordinate points; the segmented coordinate point set at multiple views and times in time is represented as... Through formula Calculating the area overlap of different detection results on the reference plane, it is not difficult to modify the calculation based on a preset overlap threshold and eliminate those that are not found. Low-confidence detection results that are less than a preset overlap threshold, among which, Indicates the test results and The ratio of coincidence on the projection plane , These are the three-dimensional coordinate point sets of the two detection results; based on the remaining high-confidence detection results, the formula is used. Determine the three-dimensional coordinate point set of the waterlogged area. Compared with area estimation results ,in, The weights are calculated by combining segment confidence, area overlap, or multi-view coverage. Let be the area of ​​the i-th detection region.

8. The method according to claim 1, wherein, The process involves projecting the three-dimensional coordinate points of the waterlogged area onto a three-dimensional overview of the foundation pit, and combining the center position of the waterlogged area's boundary box with the image frame index to generate a hazard monitoring report, including: Based on the 3D coordinate point set of the water accumulation area, the mesh structure of the 3D overview map of the foundation pit, the normal vectors of the facets, and their positional relationships, the nearest neighbor surface area corresponding to each water accumulation point is determined. Then, using vertical projection, normal projection, or shortest distance matching strategies, the projection position of the water accumulation area in the 3D overview map of the foundation pit is determined, and the 3D coordinate point set of the water accumulation area is projected onto the 3D overview map of the foundation pit. Based on the projection position of the water accumulation area in the 3D overview map of the foundation pit, the number of water accumulation targets within the current task cycle is determined, and a unique identifier is assigned to each target. Based on the center position of the bounding box of the 3D coordinate point set of the water accumulation area, the 3D position coordinates of the water accumulation area in the world coordinate system are determined, and the original image frame index corresponding to the water accumulation area is obtained. The image cropping region containing the segmentation mask of the water accumulation area is extracted as image evidence. Based on the number, number, 3D position coordinates, area, and corresponding image evidence of the water accumulation areas, a hidden danger monitoring report for the foundation pit area is generated and stored in the foundation pit area monitoring database.

9. A drone inspection system for monitoring water accumulation in foundation pits, implemented based on a drone inspection method for monitoring water accumulation in foundation pits as described in any one of claims 1-8, characterized in that, The system includes the following modules: The optimal flight monitoring route generation module is used to construct an initial route set, determine the optimal route node sequence, and generate a UAV control command set based on the spatial boundary coordinates of the pit area, flight operation altitude, camera field of view and image resolution parameters. The multi-source sensor calibration and timing alignment module is used to obtain the intrinsic and extrinsic parameters of the multi-source sensors based on the camera calibration board image, radar ranging residual and IMU motion simulation data, and to perform time alignment compensation between image frames and sensor data. The water accumulation area identification module is used to extract the segmentation mask, corresponding bounding box information and confidence score of the water accumulation area in the foundation pit area image based on historical foundation pit area image frames and corresponding timestamp information, so as to obtain the image coordinate set and attribute data of the water accumulation area. The 3D projected area estimation module is used to perform world coordinate system transformation based on the pixel mask coordinates of the image of the foundation pit area and the camera intrinsic parameters, combined with the rotation matrix constructed by Euler angles and the UAV track position vector, and to calculate the true coverage area of ​​the water accumulation area. The foundation pit 3D overview map construction module is used to extract local feature points of SIFT or ORB using multi-view images of the foundation pit area, construct cross-view correspondence, and combine the camera attitude matrix, track position vector and dense point cloud surface fitting results to generate a foundation pit 3D overview map. The multi-view water accumulation point set fusion module is used to determine the three-dimensional coordinate point set and area estimation result of the water accumulation area based on the set of three-dimensional coordinate points corresponding to the water accumulation area in each image frame, combined with the area overlap of the cross-temporal multi-view segmentation point set. The hazard monitoring report generation module is used to generate a hazard monitoring report by projecting the three-dimensional coordinate point set of the water accumulation area onto the three-dimensional overview map of the foundation pit, and combining the center position of the boundary box of the water accumulation area with the image frame index.

Images