Unmanned aerial vehicle building monitoring system and method based on binocular vision and deep learning

Abstract

The invention discloses an unmanned aerial vehicle building monitoring system and method based on binocular vision and deep learning, and the system is characterized in that an unmanned aerial vehicle carries a flight control system, an image collection module and a wireless communication module; the ground computing equipment comprises a task planning module, an image recognition and three-dimensional reconstruction module and a displacement analysis and alarm module; after the standardized visual pattern is arranged, the task planning module automatically generates a flight route according to the position of the pattern, the unmanned aerial vehicle executes the flight route and collects a synchronous image, the image is sent into the deep learning model to recognize the pattern, a three-dimensional coordinate is reconstructed, displacement analysis is performed after comparison with an initial state, and an alarm is triggered. According to the invention, the unmanned aerial vehicle platform carries the high-resolution camera for dynamic acquisition, and the deep learning algorithm is used for analysis, so that the depth perception capability of binocular stereo vision, the dynamic flexibility of the unmanned aerial vehicle platform and the environment robustness of deep learning are fused, and non-contact, full-dimension and high-economy building deformation monitoring is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of building deformation monitoring, in particular to a UAV building monitoring system and method based on binocular vision and deep learning. BACKGROUND

[0002] Existing monitoring technologies have significant limitations in the field of building deformation monitoring, making it difficult to meet the needs of non-contact, high-precision, and low-cost monitoring in complex scenarios.

[0003] Existing infrared thermographic-based outer wall shedding detection technology, such as the method disclosed in Chinese patent CN115601331A, uses the thermal conductivity difference of high-rise outer wall materials to identify outer wall shedding through infrared thermal imaging principles. However, this method has the following drawbacks: for outer wall materials with low thermal conductivity efficiency or similar thermal properties to surrounding materials, the temperature difference signal is weak, leading to detection failure; the detection effect is highly dependent on the diurnal temperature difference or specific weather conditions, and cannot work normally under weather influence; it can only provide qualitative judgment of thermal anomaly areas and cannot directly correlate material displacement or structural deformation degree, making it difficult to quantitatively assess shedding risk.

[0004] Existing crack identification technology based on UAV and deep learning uses UAV to carry two-dimensional image acquisition equipment and combines deep learning algorithms to extract crack length and width parameters. However, its limitations include: only surface two-dimensional information of cracks can be obtained, and crack depth and three-dimensional expansion trend cannot be accurately measured; it cannot comprehensively judge local or overall instability risk through dynamic evolution of cracks, making it difficult to achieve early warning before collapse, etc.

[0005] Existing digital image correlation (DIC) technology is achieved through static observation by fixed shooting at close range with high-resolution cameras. Although DIC technology can provide high-precision deformation measurement, it has the disadvantages of small monitoring range, high hardware installation cost, and dependence on fixed high-resolution camera static observation mode, requiring close-range deployment of equipment, making it difficult to cover full-scale monitoring of large or irregular structures (such as bridges and chimneys).

[0006] In summary, existing technologies have an irreconcilable contradiction between dimension coverage (only two-dimensional or thermal imaging parameters), cost, and timeliness (high-cost equipment and dependence on fixed deployment), making it difficult to meet the comprehensive needs of building deformation monitoring for full-dimensional data acquisition, dynamic real-time performance, and economy. SUMMARY

[0007] The application aims to provide a UAV building monitoring system and method based on binocular vision and deep learning, which integrates the depth perception ability of binocular stereo vision, the dynamic flexibility of the UAV platform and the environmental robustness of deep learning by means of dynamic collection by the UAV platform carrying a high-resolution camera, analysis by a deep learning algorithm, breaks through the impossible triangle of "dimension-cost-time efficiency" of the prior art, and realizes non-contact, full-dimension and high-economic building deformation monitoring.

[0008] To achieve the above-mentioned object, the application provides the following solutions.

[0009] A UAV building monitoring system based on binocular vision and deep learning, comprising a UAV and a ground computing device.

[0010] The UAV is equipped with a flight control system, an image acquisition module and a wireless communication module; the UAV performs wireless communication with the ground computing device through the wireless communication module.

[0011] The image acquisition module is a binocular camera system, comprising two physical cameras and having a synchronous triggering function; the image acquisition module is used to acquire a standardized visual pattern arranged on the surface of a building to be monitored, obtain image data, and upload the image data to the ground computing device through the wireless communication module; wherein the standardized visual pattern has a unique identification code and is arranged on the selected key stress area of the surface of the building to be monitored.

[0012] The ground computing device comprises a task planning module, an image recognition and three-dimensional reconstruction module and a displacement analysis and alarm module.

[0013] The task planning module is used to automatically calculate optimal flight waypoints and routes by clustering division and path optimization algorithm according to the spatial distribution information of the standardized visual pattern arranged on the surface of the building to be monitored and the flight characteristics of the UAV, generate a flight task instruction set, and issue the flight task instruction set to the flight control system of the UAV through the wireless communication module.

[0014] The image recognition and three-dimensional reconstruction module is used to identify the standardized visual pattern in the image data by using a deep learning target detection model, extract the center point coordinates, perform parallax calculation, combine the binocular camera calibration parameters, and reconstruct the three-dimensional space coordinates of the center point of the standardized visual pattern.

[0015] The displacement analysis and alarm module is used to compare the current three-dimensional space coordinates of the center point of the standardized visual pattern with the initial reference coordinates, calculate a three-dimensional displacement vector, and trigger an alarm and record alarm information if the preset alarm condition is met.

[0016] Further, the ground computing device further comprises a data management and result reporting module.

[0017] The data management and result reporting module is used to archive and manage all flight mission instruction sets, image data, reconstruction data and alarm information, and supports trend analysis, time series statistics, difference interpolation and three-dimensional model visualization of historical deformation data. It is also used to generate structural monitoring reports and supports local export and cloud synchronization, making it convenient for engineering and technical personnel to remotely view and connect to the smart construction platform or digital twin system.

[0018] Furthermore, the standardized visual pattern is an AprilTag or ArUco code or a customized high-contrast QR code, and the layout position of the standardized visual pattern covers the selected key stress-bearing areas on the surface of the building to be monitored, and the selected key stress-bearing areas include the center of the wall, the edge of the column, and the node connection.

[0019] Furthermore, the UAV flight characteristics include minimum flight altitude, field of view angle, and image resolution; and the flight mission instruction set is a flight control instruction that supports MAVLink or KML protocol.

[0020] Furthermore, the image data includes an intrinsic parameter, an extrinsic parameter and a baseline length parameter obtained by calibrating an image pair captured by two physical cameras.

[0021] Furthermore, the image recognition and 3D reconstruction module performs parallax calculation and, combined with the binocular camera calibration parameters, reconstructs the 3D spatial coordinates of the center point of the standardized visual pattern, specifically using the following method:

[0022] Based on the recognized standardized visual pattern, disparity calculation is performed through a stereo matching algorithm or a deep learning feature matching algorithm. Combined with the binocular camera calibration parameters, the 3D spatial coordinates of the center point of the standardized visual pattern are reconstructed. The reconstruction formula is as follows:

[0023]

[0024] in, is the camera focal length, is the binocular baseline length, is the disparity value; is the coordinate of the principal point; x and y are the horizontal and vertical pixel positions of the center point of the standardized visual pattern on the two-dimensional plane of the image, respectively;

[0025] The image recognition and 3D reconstruction module outputs the 3D spatial coordinates (X, Y, Z) of the center point of each standardized visual pattern.

[0026] Furthermore, the initial reference coordinates are the three-dimensional spatial coordinates of each standardized visual pattern recorded during the first flight; the alarm information includes the pattern number, alarm time, three-dimensional displacement limit value and direction.

[0027] Further, the preset alarm condition is that a displacement vector in any coordinate axis direction exceeds a preset threshold, a module length of a three-dimensional displacement vector is out of limit, or a multi-condition combination alarm rule defined by a user.

[0028] The application further provides a UAV building monitoring method based on binocular vision and deep learning, applied to the UAV building monitoring system based on binocular vision and deep learning.

[0029] S1, a standard visual pattern with a unique identification code is arranged on a key area on a surface of a building to be monitored;

[0030] S2, according to spatial distribution information of the standard visual pattern arranged on the surface of the building to be monitored and flight characteristics of a UAV, an optimal flight waypoint and a flight route are automatically calculated through a clustering division and a path optimization algorithm, a flight task instruction set is generated, and is sent to a flight control system of the UAV;

[0031] S3, the UAV flies along the flight route to each waypoint and hovers, a left-right image pair of the standard visual pattern arranged on the surface of the building to be monitored is collected through a synchronously triggered binocular camera system, image data is obtained, and is uploaded to a ground computing device;

[0032] S4, the ground computing device adopts a deep learning target detection model to identify the standard visual pattern in the image data, extracts a center point coordinate, and performs parallax calculation, combines binocular camera calibration parameters, and reconstructs three-dimensional space coordinates of the center point of the standard visual pattern;

[0033] S5, the current three-dimensional space coordinates of the center point of the standard visual pattern are compared with initial reference coordinates, a three-dimensional displacement vector is calculated, if a preset alarm condition is met, an alarm is triggered and alarm information is recorded.

[0034] Further, the method further comprises:

[0035] S6, monitoring data is archived and a visual analysis report is generated.

[0036] Compared with the prior art, the UAV building monitoring system and method based on binocular vision and deep learning provided by the application first systematically integrate binocular vision imaging, deep learning identification, three-dimensional reconstruction calculation, UAV automatic navigation and building deformation monitoring, and build a building facade micro-deformation monitoring system with high precision, low cost, non-contact and repeatable automatic execution.

[0037] The application utilizes the depth perception capability of the binocular camera to construct three-dimensional point cloud data of the building surface, and can realize accurate measurement of crack depth, displacement and deformation trend; through flexible flight path planning of the unmanned aerial vehicle, the monitoring range and observation point limitation are broken through, and the application is suitable for scenes such as bridges, chimneys and other fixed monitoring points; combined with a high-robustness deep learning algorithm, the standard visual pattern (AprilTag or ArUco) arranged on the building surface is automatically identified and positioned, which greatly improves the stability, accuracy and environmental adaptability (such as strong light, inclination and shielding) of pattern recognition; the application can obtain two-dimensional topography and three-dimensional deformation data in real time without contacting the building surface, and provides quantitative basis for structure risk assessment.

[0038] In summary, through technical fusion, the application realizes the coordinated optimization of building deformation monitoring in the three dimensions of "dimension-cost-time efficiency", and provides a high economic and adaptive solution for building safety assessment, which is suitable for various building structures, including high-rise buildings, bridges, factories, ancient buildings and the like, and only needs to arrange standard monitoring patterns (such as AprilTag, ArUco labels and the like) on the outer wall of the monitoring area, so that centimeter-level deformation recognition and monitoring can be realized. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed to be used in the embodiments will be briefly introduced. Obviously, the drawings in the following description only constitute some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0040] Figure 1 FIG. 1 is a structural schematic diagram of an unmanned aerial vehicle building monitoring system based on binocular vision and deep learning according to the application, wherein (a) is an unmanned aerial vehicle, and (b) is a building to be measured;

[0041] Figure 2 FIG. 2 is a building deformation judgment schematic diagram;

[0042] Figure 3 FIG. 3 is a method schematic diagram of an unmanned aerial vehicle building monitoring method based on binocular vision and deep learning according to the application;

[0043] The specific embodiments of the application will be described below with reference to the drawings. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0045] The purpose of this invention is to provide a UAV building monitoring system and method based on binocular vision and deep learning, which combines the UAV flight platform, binocular stereo vision and deep learning technology for non-contact detection of tiny three-dimensional deformation (centimeter level) of building structure surfaces.

[0046] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] Example 1

[0048] like Figure 1 and Figure 2 As shown, the UAV building monitoring system based on binocular vision and deep learning provided by the present invention includes: a UAV and a ground computing device.

[0049] The UAV is equipped with a flight control system, an image acquisition module and a wireless communication module; the UAV communicates wirelessly with a ground computing device via the wireless communication module.

[0050] The image acquisition module is a binocular camera system, which includes two physical cameras on the left and right and has a synchronous triggering function; the image acquisition module is used to collect standardized visual patterns arranged on the surface of the building to be monitored, obtain image data, and upload it to the ground computing device through the wireless communication module; wherein, the standardized visual pattern has a unique identification code and is arranged on the selected key stress-bearing areas on the surface of the building to be monitored; the standardized visual pattern is an AprilTag or ArUco code or a customized high-contrast QR code, which has unique identification, high robustness and stable geometric features, and is used for subsequent image recognition and spatial positioning. The pattern layout should take into account the building scale, monitoring accuracy requirements and structural importance to ensure coverage of key stress-bearing areas (such as the center of the wall, the edge of the column, the node connection, etc.).

[0051] The drone automatically flies according to the flight mission instruction set, hovering at each waypoint and synchronously capturing left and right images. The image pairs captured by the binocular camera system are calibrated to obtain intrinsic and extrinsic parameters for baseline length to ensure the accuracy of subsequent stereo matching and 3D reconstruction.

[0052] The ground computing device comprises a task planning module, an image recognition and three-dimensional reconstruction module, a displacement analysis and alarm module, and a data management and result reporting module.

[0053] The task planning module is used to automatically calculate optimal flight waypoints and routes according to the spatial distribution information of the standardized visual patterns laid on the surface of the building to be monitored and the flight characteristics of the unmanned aerial vehicle (such as the minimum flight height, the field of view angle, the image resolution, and other parameters), generate a flight task instruction set, and issue the flight task instruction set to the flight control system of the unmanned aerial vehicle through the wireless communication module. The flight task instruction set ensures that the unmanned aerial vehicle efficiently covers all the standardized visual patterns in the shortest time, and the generated task file supports common flight control protocols (such as MAVLink, KML, etc.), which can be directly used for the flight control system.

[0054] The image recognition and three-dimensional reconstruction module is the core module of the system, and mainly comprises the following two functions:

[0055] (1) Pattern recognition: a deep learning detection model (such as YOLOv5, DETR) is used to automatically recognize the monitoring patterns in the image and extract the center point coordinates thereof.

[0056] (2) Three-dimensional reconstruction: as shown in Figure 2 , based on a binocular image pair, a traditional stereo matching algorithm (such as semi-global block matching (SGBM)) or a deep learning feature matching algorithm (such as SuperGlue+LoFTR) is used for disparity calculation, and three-dimensional coordinate restoration is performed in combination with the camera parameters. The reconstruction formula is as follows:

[0057]

[0058] wherein, is the focal length of the camera, is the binocular baseline length, is the disparity value; is the principal point coordinate; x and y are respectively the horizontal and vertical pixel positions of the center point of the standardized visual pattern in the two-dimensional plane of the image; the principal point coordinate is the representative spatial position of the monitoring point pattern, which is called in the subsequent feature matching, three-dimensional reconstruction, and deformation calculation modules, and is used to realize accurate position correspondence and space deformation calculation;

[0059] The image recognition and three-dimensional reconstruction module outputs the three-dimensional spatial coordinates (X, Y, Z) of the center point of each standardized visual pattern.

[0060] The displacement analysis and alarm module is used for comparing the current three-dimensional space coordinates of the center point of the standardized visual pattern with the initial reference coordinates, calculating a three-dimensional displacement vector, triggering an alarm and recording alarm information if a preset alarm condition is met; wherein the initial reference coordinates are the three-dimensional space coordinates of each standardized visual pattern recorded during the first flight.

[0061] Calculate the three-dimensional displacement vector of each standardized visual pattern :

[0062]

[0063] wherein, , , ) is the initial reference coordinate.

[0064] The preset alarm condition supports three kinds of judgment strategies:

[0065] 1. Any direction , , exceeds the preset threshold value;

[0066] 2. Three-dimensional displacement modulus (Euclidean distance) is over limit;

[0067] 3. User-defined multi-condition combination alarm rule.

[0068] The system records the alarm information (pattern number, alarm time, over-limit value and direction) and can push it to the terminal or early warning equipment.

[0069] The data management and result reporting module is used for archiving and managing all flight task instruction sets, image data, reconstruction data and alarm information, and supports trend analysis, time series statistics, difference interpolation and three-dimensional model visualization of historical deformation data, and is used for generating a structure monitoring report and supporting local and cloud synchronization export, facilitating remote viewing and interfacing with a smart construction platform or a digital twin system by engineering and technical personnel.

[0070] The working principle of the unmanned aerial vehicle building monitoring system based on binocular vision and deep learning provided by the application is as follows: after the pattern is laid out, the task planning module automatically generates a flight route according to the pattern position, the unmanned aerial vehicle executes the flight route and collects synchronous images, the images are sent to a deep learning model to identify the pattern, a three-dimensional coordinate is reconstructed through a stereo matching algorithm, displacement analysis is performed after comparison with the initial state, and an alarm is triggered.

[0071] The unmanned aerial vehicle building monitoring system based on binocular vision and deep learning provided by the application realizes the following technical effects:

[0072] 1. System integration innovation: The invention first integrates binocular vision imaging, deep learning recognition, three-dimensional reconstruction calculation, unmanned aerial vehicle automatic navigation, and building deformation monitoring, and builds a high-precision, low-cost, non-contact, and repeatable automatic building facade micro-deformation monitoring system.

[0073] 2. Pattern recognition mechanism based on deep learning: The deep learning algorithm (such as YOLOv5) is used to automatically identify and locate the standard visual patterns (AprilTag or ArUco) laid on the building surface, greatly improving the stability, accuracy, and environmental adaptability of pattern recognition (such as strong light, inclination, and occlusion).

[0074] 3. Use binocular vision to realize centimeter-level three-dimensional deformation measurement: Through binocular camera + stereo matching algorithm, the accurate coordinates of the monitoring points in three-dimensional space can be obtained without manual calibration, realizing the extraction and analysis of the centimeter-level three-dimensional displacement vector (X, Y, Z) of the outer wall pattern. 、 、

[0075] 4. Unmanned aerial vehicle route automatic planning and multi-point task execution mechanism: Combined with the building structure diagram and pattern layout, the system can automatically generate flight routes, set flight point positions, adjust flight attitudes and dwell times, and repeatedly execute flight tasks to compare and analyze the periodic structure state and trends.

[0076] 5. Multi-module data fusion type deformation judgment and alarm mechanism: Fusion of deep recognition, image three-dimensional reconstruction, and time series data changes forms a complete data processing chain, realizes automatic recognition, judgment, comparison, alarm, and result report generation of structure deformation, and can be used for real-time monitoring and historical comparison.

[0077] 6. Wide range of applications: Not dependent on building type and material characteristics, as long as the outer wall can be pasted with patterns, the system can be applied, suitable for high-rise buildings, bridges, industrial plants, ancient buildings, tunnel entrances, etc.

[0078] 7. Highly modular system structure: The hardware and software of the system have the characteristics of modularity, replaceability, and scalability, such as deep model, pattern design, flight control system, and data processing module can be freely upgraded, with good adaptability and maintainability.

[0079] Embodiment 2

[0080] As shown in Figure 3 , the invention also provides an unmanned aerial vehicle building monitoring method based on binocular vision and deep learning, applied to the above-mentioned unmanned aerial vehicle building monitoring system based on binocular vision and deep learning, comprising the following steps:

[0081] ​S1, pattern placement: Place standardized visual patterns with unique identification codes on selected key areas of the surface of the building to be tested; S2, Mission Planning: Based on the spatial distribution of standardized visual patterns on the surface of the monitored building and the flight characteristics of the drone, the optimal flight waypoints and routes are automatically calculated through clustering and path optimization algorithms, and a flight mission instruction set is generated and sent to the drone's flight control system; S3, UAV acquisition: The UAV flies along the route to each waypoint and hovers to perform monitoring and acquisition. Specifically, a synchronously triggered binocular camera system collects left and right image pairs of the standardized visual pattern arranged on the surface of the monitored building, obtains image data, and uploads it to the ground computing device; S4, Pattern Recognition and 3D Reconstruction: Ground computing equipment uses a deep learning target detection model to identify standardized visual patterns in image data, extract the coordinates of their center points, and perform disparity calculations. Combined with the calibration parameters of the binocular camera, the 3D spatial coordinates of the center points of the standardized visual patterns are reconstructed to achieve 3D reconstruction. S5, displacement analysis and alarm: Compare the current 3D spatial coordinates of the center point of the standardized visual pattern with the initial reference coordinates, calculate the 3D displacement vector, and trigger an alarm and record the alarm information if the preset alarm conditions are met. The initial reference coordinates are the 3D spatial coordinates of each standardized visual pattern first collected during the first flight of the UAV after completing the mission planning in step S2. S6, Data Management and Result Reporting: Archive monitoring data and generate visual analysis reports.

[0082] During the execution of the UAV building monitoring method based on binocular vision and deep learning:

[0083] The user sets the mission parameters and arranges the monitoring pattern through the ground station;

[0084] The system automatically generates a flight path based on the parameters;

[0085] Start the flight, complete waypoint shooting and send back images;

[0086] The system automatically analyzes and generates monitoring reports;

[0087] If the deformation exceeds the limit, the system will push an alarm message to the user terminal.

[0088] The system supports repeated flights and periodic inspections and can be used as a long-term health monitoring tool for buildings.

[0089] The UAV building monitoring method based on binocular vision and deep learning provided by the present invention achieves the following technical effects:

[0090] 1. Overcome the shortcomings of infrared thermal imaging technology and move from passive lag to active early warning.

[0091] The prior art infrared thermal imaging detection relies on material temperature difference, and cannot form thermal anomaly in the initial stage of hollowing (deformation <1 cm), resulting in high missed detection rate.

[0092] The present application has a direct displacement measurement mechanism, which realizes a depth resolution of ±1.2 mm at a distance of 10 meters through binocular stereo vision (baseline 0.5 m) and SuperGlue sub-pixel matching (error ±0.2 pixels), can accurately capture 0.5 cm level micro-deformation (such as wall bulge), and has damage precursor identification capability, displacement vector calculation directly quantifies structure deformation, and does not need to wait for heat conduction effect.

[0093] II. Breakthrough the limitations of crack identification technology, from two-dimensional appearance to three-dimensional essence.

[0094] In the prior art, the traditional crack detection can only obtain length / width parameters, and cannot evaluate the expansion in the depth direction (such as crack deepening caused by internal steel bar corrosion).

[0095] The present application can realize three-dimensional strain field analysis, and the binocular SGBM algorithm generates 20,000 points / m 2 dense point cloud, which can also be combined with spatial strain formula to calculate the Z-direction strain of the crack edge and quantify the displacement in the depth direction.

[0096] The present application can also analyze crack area displacement time series data (such as ΔZ daily average increment >0.5 mm) through LSTM network to predict internal structure instability probability.

[0097] III. Subvert the DIC technology paradigm, from static high cost to dynamic economy.

[0098] In the prior art, the digital image correlation method (DIC) needs to fixedly deploy high-resolution cameras (single camera covers <10 m 2 ), which is too high for high-rise building monitoring, and cannot adapt to curved structures (such as chimneys).

[0099] The present application has dynamic coverage capability, and the unmanned aerial vehicle platform cooperates with RTK centimeter-level positioning (horizontal ±1 cm), and can scan 200 m 2 facade in 30 minutes per flight, which is 12 times the efficiency of fixed DIC; the present application adopts adaptive flight path control algorithm, which can dynamically encrypt flight points according to |ΔZ| value (from 1 point / m 2 →10 points / m 2 ), replacing dozens of fixed cameras. In curved structure monitoring, image distortion is solved by 45° oblique angle supplementary shooting.

[0100] The present application has other known technologies.

[0101] The principles and implementation manners of the present application are described by using specific examples in the present application, and the above examples are only used to help understand the method of the present application and its core idea; meanwhile, for the general technical personnel in the art, the specific implementation manners and application ranges will be changed according to the idea of the present application. In conclusion, the content of the present specification should not be understood as the limitation of the present application.

Claims

1. A UAV building monitoring system based on binocular vision and deep learning, characterized by: include: drones and ground computing equipment; The UAV is equipped with a flight control system, an image acquisition module and a wireless communication module; the UAV communicates wirelessly with the ground computing device via the wireless communication module; The image acquisition module is a binocular camera system consisting of two physical cameras, one on the left and one on the right, with a synchronous triggering function. The image acquisition module is used to capture standardized visual patterns on the surface of the building to be monitored, obtain image data, and upload it to the ground computing device via a wireless communication module. The standardized visual patterns have a unique identification code and are placed on selected key stress-bearing areas on the surface of the building to be monitored. The ground computing equipment includes a mission planning module, an image recognition and 3D reconstruction module, and a displacement analysis and alarm module; The mission planning module is used to automatically calculate the optimal flight waypoints and routes based on the spatial distribution information of the standardized visual patterns on the surface of the monitored building and the flight characteristics of the drone through clustering and path optimization algorithms, generate a flight mission instruction set, and send it to the drone's flight control system through the wireless communication module; The image recognition and 3D reconstruction module is used to use a deep learning target detection model to identify standardized visual patterns in image data, extract the coordinates of their center points, and perform disparity calculations. Combined with the binocular camera calibration parameters, the 3D spatial coordinates of the center points of the standardized visual patterns are reconstructed. The displacement analysis and alarm module is used to compare the current three-dimensional space coordinates of the center point of the standardized visual pattern with the initial reference coordinates, calculate the three-dimensional displacement vector, and trigger an alarm and record the alarm information if the preset alarm conditions are met.

2. The UAV building monitoring system based on binocular vision and deep learning according to claim 1 is characterized in that: The ground computing equipment also includes a data management and result reporting module; The data management and result reporting module is used to archive and manage all flight mission instruction sets, image data, reconstruction data and alarm information, and supports trend analysis, time series statistics, difference interpolation and three-dimensional model visualization of historical deformation data. It is also used to generate structural monitoring reports and supports local export and cloud synchronization, making it convenient for engineering and technical personnel to remotely view and connect to the smart construction platform or digital twin system.

3. The UAV building monitoring system based on binocular vision and deep learning according to claim 1 is characterized in that: The standardized visual pattern is an AprilTag or ArUco code or a customized high-contrast QR code. The layout position of the standardized visual pattern covers the selected key stress-bearing areas on the surface of the building to be monitored. The selected key stress-bearing areas include the center of the wall, the edge of the column, and the node connection.

4. The UAV building monitoring system based on binocular vision and deep learning according to claim 1 is characterized in that: The UAV flight characteristics include minimum flight altitude, field of view angle, and image resolution; the flight mission instruction set is a flight control instruction that supports the MAVLink or KML protocol.

5. The UAV building monitoring system based on binocular vision and deep learning according to claim 1 is characterized in that: The image data includes an internal reference and an external reference baseline length parameter obtained by calibrating an image pair captured by two physical cameras.

6. The UAV building monitoring system based on binocular vision and deep learning according to claim 1 is characterized in that: The image recognition and 3D reconstruction module performs parallax calculation and, combined with the binocular camera calibration parameters, reconstructs the 3D spatial coordinates of the center point of the standardized visual pattern. Specifically, the following method is used: Based on the recognized standardized visual pattern, disparity calculation is performed through a stereo matching algorithm or a deep learning feature matching algorithm. Combined with the binocular camera calibration parameters, the 3D spatial coordinates of the center point of the standardized visual pattern are reconstructed. The reconstruction formula is as follows: ； in, is the camera focal length, is the binocular baseline length, is the disparity value; is the coordinate of the principal point; x and y are the horizontal and vertical pixel positions of the center point of the standardized visual pattern on the two-dimensional plane of the image, respectively; The image recognition and 3D reconstruction module outputs the 3D spatial coordinates (X, Y, Z) of the center point of each standardized visual pattern.

7. The UAV building monitoring system based on binocular vision and deep learning according to claim 1 is characterized in that: The initial reference coordinates are the three-dimensional spatial coordinates of each standardized visual pattern recorded during the first flight; the alarm information includes the pattern number, alarm time, three-dimensional displacement limit value and direction.

8. The UAV building monitoring system based on binocular vision and deep learning according to claim 1 is characterized in that: The preset alarm conditions are: the displacement vector in any coordinate axis direction exceeds the preset threshold, the modulus of the three-dimensional displacement vector exceeds the limit, or a user-defined multi-condition combination alarm rule.

9. A UAV building monitoring method based on binocular vision and deep learning, applied to the UAV building monitoring system based on binocular vision and deep learning according to any one of claims 1 to 8, characterized in that: The steps include: S1, laying out standardized visual patterns with unique identification codes on selected key areas of the surface of the building to be tested; S2, based on the spatial distribution of standardized visual patterns on the surface of the monitored building and the flight characteristics of the drone, automatically calculates the optimal flight waypoints and routes through clustering and path optimization algorithms, generates a flight mission instruction set, and sends it to the drone's flight control system; In step S3, the UAV flies to each waypoint according to the route and hovers. The UAV uses a synchronously triggered binocular camera system to collect left and right image pairs of the standardized visual pattern arranged on the surface of the monitored building, obtains image data, and uploads it to the ground computing device; S4: The ground computing device uses a deep learning target detection model to identify the standardized visual pattern in the image data, extracts the coordinates of its center point, and performs disparity calculation. Combined with the binocular camera calibration parameters, it reconstructs the 3D spatial coordinates of the center point of the standardized visual pattern. S5, compare the current three-dimensional space coordinates of the center point of the standardized visual pattern with the initial reference coordinates, calculate the three-dimensional displacement vector, and if the preset alarm conditions are met, trigger an alarm and record the alarm information.

10. The UAV building monitoring method based on binocular vision and deep learning according to claim 9 is characterized in that: The method further comprises: S6, archives monitoring data and generates visual analysis reports.

Images