Dynamic monitoring method and system for high slope based on visual detection

Abstract

This invention relates to the field of visual inspection technology, and discloses a dynamic monitoring method and system for high slopes based on visual inspection. The invention predicts current abnormal areas based on multiple images and the overall morphology of the high slope. It then triggers secondary imaging by ground cameras and drones based on these abnormal areas to determine local images of the current abnormal areas and identify corresponding abnormal events, thus improving the initial accuracy of abnormal event detection. Based on the detection of these abnormal events, multiple abnormal features are determined. The anomaly coefficient of each anomaly event is determined based on its feature location, corresponding feature shape, and past anomalies, and the anomaly level of the event is marked. If there are multiple anomalies, an anomaly control system for high slopes is constructed based on the anomaly levels of these multiple anomalies, their corresponding abnormal areas, and the overall morphology of the high slope, improving the accuracy of the high slope anomaly control system.

Description

Technical Field

[0001] This invention relates to the technical field of visual inspection, and more particularly to a dynamic monitoring method and system for high slopes based on visual inspection. Background Technology

[0002] With the development of technology, high slopes, as a type of slope, belong to architectural features. High slopes usually refer to those formed during engineering construction (such as the construction of highways, railways, dams, mining, or urban development) or existing in a natural state, with a relatively large height (usually referring to soil slopes with a height greater than 20 meters and rock slopes with a height greater than 30 meters). In the current technology, high slopes are visually inspected based on cameras, and multiple images of the high slopes are collected. The corresponding abnormal areas are determined based on the recognition of multiple images. However, the secondary shooting by ground cameras and drones is ignored, which affects the initial accuracy of abnormal events and leads to the low accuracy of the high slope anomaly control system. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a dynamic monitoring method and system for high slopes based on visual detection.

[0004] This invention provides a dynamic monitoring method for high slopes based on visual detection, comprising: Mark the current location of the high slope, determine the monitoring space of the high slope based on the current location of the high slope, the corresponding overall shape and previous monitoring events, and divide the high slope into multiple sub-monitoring spaces based on visual detection of the monitoring space; The multi-layered sub-monitoring spaces are matched with ground cameras and drones to construct a multi-dimensional monitoring system for high slopes. In the multi-dimensional monitoring system, ground cameras perform the first layer of visual inspection on the sub-monitoring spaces, and drones perform the second layer of visual inspection on the remaining sub-monitoring spaces. The multi-dimensional monitoring system specifies the responsible platform, monitoring frequency, and data quality requirements for each sub-space. Based on the first and second levels of monitoring, multiple images are determined. Based on these multiple images and the overall shape of the high slope, the current abnormal area is predicted. The current abnormal area is then used to trigger secondary shooting by ground cameras and drones to determine local images of the current abnormal area and identify the corresponding abnormal event. The abnormal event includes event ID, type, location, geometric parameters, risk level, and discovery time. Based on the detection of the abnormal event, multiple abnormal features are identified. Based on the feature location, corresponding feature shape and previous abnormal events of the multiple abnormal features, the abnormal coefficient of the abnormal event is determined, and the abnormal level of the abnormal event is marked. If there are multiple abnormal events, an abnormal control system for the high slope is constructed based on the abnormality level of the multiple abnormal events, the corresponding abnormal areas, and the overall shape of the high slope, so as to trigger the unscheduled inspection path of the drone and mark the key detection areas of the ground camera.

[0005] This invention provides a dynamic monitoring system for high slopes based on visual detection, which is applied to the aforementioned dynamic monitoring method for high slopes based on visual detection.

[0006] Compared with the prior art, the beneficial effects of the present invention are: (1) Mark the current location of the high slope, determine the monitoring space of the high slope based on the current location of the high slope, the corresponding overall shape and previous monitoring events, and divide the high slope into multiple sub-monitoring spaces based on the visual detection of the monitoring space of the high slope; match the multiple sub-monitoring spaces to ground cameras and drones, and construct a multi-dimensional monitoring system for the high slope. In the multi-dimensional monitoring system, the ground camera performs the first visual detection of the sub-monitoring space, and the drone performs the second visual detection of the remaining sub-monitoring space; determine multiple images based on the first and second monitoring, predict the current abnormal area based on the multiple images and the overall shape of the high slope, and trigger the ground camera and drone to take a second shot based on the current abnormal area to determine the local image of the current abnormal area and determine the corresponding abnormal event. The multi-dimensional monitoring system for the high slope is introduced to further control the first and second visual detection, and is compatible with the local image of the current abnormal area, thus improving the initial accuracy of the abnormal event.

[0007] (2) Based on the detection of the abnormal event, multiple abnormal features are determined. Based on the feature location, corresponding feature shape and previous abnormal events of the multiple abnormal features, the abnormal coefficient of the abnormal event is determined and the abnormal level of the abnormal event is marked. If there are multiple abnormal events, the abnormal control system of the high slope is constructed based on the abnormal level of multiple abnormal events, the corresponding abnormal area and the overall shape of the high slope, so as to trigger the unscheduled inspection path of the UAV and mark the key detection area of ​​the ground camera. The abnormal level of the abnormal event is introduced, realizing the overall consideration of the abnormal level of multiple abnormal events, the corresponding abnormal area and the overall shape of the high slope, improving the accuracy of the abnormal control system of the high slope, and triggering the unscheduled inspection path of the UAV and marking the key detection area of ​​the ground camera. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the dynamic monitoring method for high slopes based on visual detection in an embodiment of the present invention. Figure 2This is a flowchart illustrating step S11 in the dynamic monitoring method for high slopes based on visual detection in this embodiment of the invention. Figure 3 This is a flowchart illustrating step S12 in the dynamic monitoring method for high slopes based on visual detection in this embodiment of the invention. Figure 4 This is a flowchart illustrating step S13 in the dynamic monitoring method for high slopes based on visual detection in this embodiment of the invention. Figure 5 This is a flowchart illustrating step S14 in the dynamic monitoring method for high slopes based on visual detection in this embodiment of the invention. Figure 6 This is a flowchart illustrating step S15 in the dynamic monitoring method for high slopes based on visual detection in this embodiment of the invention. Figure 7 This is a schematic diagram of the structural composition of a dynamic monitoring system for high slopes based on visual detection in an embodiment of the present invention. Detailed Implementation

[0009] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0010] Please see Figures 1 to 7 A dynamic monitoring method for high slopes based on visual inspection is proposed and applied to visual inspection scenarios. The dynamic monitoring method for high slopes based on visual inspection includes: Step S11: Mark the current location of the high slope, determine the monitoring space of the high slope based on the current location of the high slope, the corresponding overall shape and previous monitoring events, and divide the high slope into multiple sub-monitoring spaces based on visual detection of the monitoring space. Step S12: Match the multi-layer sub-monitoring space to ground cameras and drones, and construct a multi-dimensional monitoring system for high slopes. In the multi-dimensional monitoring system, the ground cameras perform the first visual inspection of the sub-monitoring space, and the drones perform the second visual inspection of the remaining sub-monitoring space. Step S13: Based on the first and second monitoring, multiple images are determined. Based on the multiple images and the overall shape of the high slope, the current abnormal area is predicted. The ground camera and drone are triggered to take secondary pictures according to the current abnormal area to determine the local image of the current abnormal area and determine the corresponding abnormal event. Step S14: Based on the detection of the abnormal event, determine multiple abnormal features, determine the abnormality coefficient of the abnormal event based on the feature location, corresponding feature shape and previous abnormal events of the multiple abnormal features, and mark the abnormality level of the abnormal event. Step S15: If there are multiple abnormal events, construct an abnormal control system for the high slope based on the abnormality level of the multiple abnormal events, the corresponding abnormal area, and the overall shape of the high slope, so as to trigger the unscheduled inspection path of the UAV and mark the key detection area of ​​the ground camera.

[0011] refer to Figure 2 In step S11, the specific steps are as follows: S111: Collect the location coordinates of the high slope and determine the current location of the high slope. Determine the overall shape of the high slope based on the map information of the high slope. Determine the monitoring space of the high slope based on the current location of the high slope, the corresponding overall shape, and previous monitoring events. S112: Visually inspect the monitoring space of high slopes and dynamically monitor the monitoring space of high slopes in different dimensions. Divide the monitoring space of high slopes into multiple spatial markers during the division process. Based on the multiple spatial markers, the location of the ground camera and the flight path of the UAV, determine multiple sub-monitoring spaces and mark the monitoring range of each sub-monitoring space.

[0012] In the embodiments of this application, the location coordinates of the high slope are collected and the current location of the high slope is determined. The overall shape of the high slope is determined based on the drawing information of the high slope. The monitoring space of the high slope is determined based on the current location of the high slope, the corresponding overall shape, and previous monitoring events. This approach takes into account the overall consideration of the current location of the high slope, the corresponding overall shape, and previous monitoring events, ensuring the accuracy of the monitoring space of the high slope.

[0013] At this point, Global Navigation Satellite System Real-Time Dynamic (GNSS-RTK) measurement technology is adopted. Several geologically stable, open-view, and easily damaged locations are selected around the high slope and on the slope itself to set up permanent monitoring and control points. Differential calculations are performed between GNSS-RTK receivers and Continuously Operating Reference Stations (CORS) or self-established reference stations to obtain the three-dimensional coordinates of these control points in a specific coordinate system. The accuracy can usually reach the centimeter or even millimeter level. The output is a series of control points with precise three-dimensional coordinates. These are not only anchor points for calibrating geographical locations, but also reference anchor points for all subsequent spatial registration and coordinate transformation of data.

[0014] Acquiring macroscopic geometric morphology and internal structural information of high slopes to construct their digital skeleton requires integrating data from both theoretical design and actual conditions. The theoretical morphology data sources include design / as-built drawings and geological survey maps. The former provides theoretical geometric dimensions and internal reinforcement structure information, while the latter reveals key internal genetic information such as geological profiles, lithological boundaries, and weak interlayers. Acquiring actual morphology data mainly relies on UAV oblique photogrammetry and ground-based 3D laser scanning (LiDAR) to generate high-precision 3D point cloud models and digital orthophotos, efficiently acquiring the real texture and morphology of the slope surface. By using the control points established in the first step as a benchmark, the point cloud model acquired on-site is precisely registered with CAD drawings to generate a refined 3D model that integrates design and current status information. This model not only shows the real morphology of the slope but also reveals its internal structure.

[0015] All previous monitoring events, damage records, and maintenance history are digitized and spatialized, and precisely marked as geographic elements on the 3D model constructed in the second step. This historical data is key prior knowledge for assessing current risks. The boundary of the monitoring space is not a simple slope body, but a 3D volume that has been risk-assessed and buffered: its upper boundary is the slope crest line extending outwards to a certain distance to capture potential trailing edge tension cracks; its lower boundary is the slope toe line extending outwards beyond the critical facilities threatened by it; and its lateral boundaries extend along the slope direction to stable ridges or geological structure boundaries. The final output is a digital monitoring space with clear 3D boundaries that integrates geological, design, current morphology, and historical risk information.

[0016] Specifically, for the high slope, technicians set up a total of 5 permanent monitoring control points (numbered P01-P05) on the stable bedrock at the top of the slope, the hardened road surface at the bottom of the slope, and on the mountains on both sides. Through multi-time GNSS-RTK static observation and calculation, the precise coordinates of these 5 points in the engineering coordinate system were determined. For example, the coordinates of point P01 (located at the top of the slope) are (X:3851234.567, Y:512345.789, H:215.345). This coordinate system consisting of P01-P05 became the benchmark for all subsequent surveying work.

[0017] The original CAD drawings and detailed geological survey report of the high slope were collected. The report indicated that there was a weak interlayer about 0.5 meters thick at a slope height of 80 meters. Aerial survey of the high slope was conducted using a drone equipped with lidar, generating a three-dimensional point cloud model containing 50 million points. The model clearly showed that due to years of rain erosion, a distinct gully had formed in the upper part of the slope, which was not present in the original design drawings. Using control points P01-P05 as a reference, the technicians accurately projected the location of the weak interlayer from the geological profile onto the three-dimensional point cloud model, ultimately obtaining a digital model that showed both the actual slope surface morphology (including the gully) and marked the location of the internal, invisible weak interlayer.

[0018] When constructing the monitoring space, the team reviewed the archives and found that during the rainy season of 2019, a shallow landslide with a volume of about 50 cubic meters occurred at an elevation of about 85 meters on the high slope (coinciding with the location of the weak interlayer). At that time, shotcrete and anchor support had been carried out. The technicians accurately delineated the boundary of the restoration area on the 3D model and marked it as a historical landslide restoration area, with a risk weight set to high. The final high slope monitoring space was determined with its upper boundary 30 meters outside the slope crest line, its lower boundary the center line of the G50 National Highway below the slope foot, and its lateral boundaries extending to the natural valleys on both sides. This 3D space not only included the slope itself but also the national highway, a key protected object, and marked the 2019 landslide restoration area as a high-risk sub-area. Thus, a complete and information-rich high slope monitoring space was successfully constructed.

[0019] Furthermore, visual inspection is performed on the monitoring space of high slopes, and the monitoring space of high slopes is dynamically monitored in different dimensions. The monitoring space of high slopes is spatially divided, and multiple spatial markers are determined during the division process. Based on multiple spatial markers, the position of ground cameras and the flight path of drones, multiple sub-monitoring spaces are determined, and the monitoring range of each sub-monitoring space is marked. This approach takes into account the overall consideration of multiple spatial markers, the position of ground cameras and the flight path of drones, ensuring the accuracy of the multi-layer sub-monitoring spaces.

[0020] At this point, based on the three-dimensional model constructed by S111, multi-dimensional and visualized analysis and evaluation are carried out to provide a basis for spatial division. Dimensions refer to different factors that affect slope stability, mainly including: geological dimension (identifying internal factors such as lithological boundaries, faults, and weak interlayers), geometric dimension (calculating slope and slope height, and identifying unfavorable terrain such as steep areas), hydrological dimension (analyzing catalyst factors such as surface runoff paths and catchment areas), and historical risk dimension (marking historical event areas). The division method can adjust the weight of different dimensions according to seasonal changes, such as automatically increasing the priority of hydrologically related areas during the rainy season.

[0021] High-risk feature points identified in multidimensional analysis are used as the initial region. The region expands outward from the initial region, and the boundary of the expansion is determined by the gradient of risk factors. When the slope, lithology, or risk level changes significantly, growth stops at this boundary. The entire monitoring space is divided into a series of irregularly shaped, differently sized, but relatively homogeneous sub-monitoring spaces. While dividing the space, the system automatically assigns a set of descriptive spatial labels to each sub-space, including a unique identifier, geometric attributes, geological attributes, risk level, and main monitoring targets.

[0022] The system will match the spatial markers of all subspaces, the deployment parameters of ground cameras (coordinates, attitude, focal length), and the performance parameters of drones (endurance, altitude). The first layer is the ground camera layer, which selects subspaces with high risk levels that can be effectively covered by cameras for high-frequency, continuous monitoring. The second layer is the drone inspection layer, which includes areas with medium or low risk levels, or high-risk areas that cannot be covered by cameras, for periodic, mobile, wide-area coverage.

[0023] The system provides the direct basis for operation; the monitoring range is defined as precise geographic information, usually using a polygon coordinate sequence; for ground cameras, the system calculates the geographic coordinates of the smallest bounding rectangle or irregular polygon that can clearly cover a specific subspace; for drones, it generates the desired aerial photography area polygon and the best waypoint coordinates for each subspace; finally, it outputs a structured list of subspaces, with each entry clearly defined as: [subspace ID, level, boundary coordinate sequence, responsible device ID, monitoring frequency requirement].

[0024] Specifically, a multi-dimensional analysis was initiated on the 3D model of the high slope: Geologically, the system automatically highlighted the weak interlayer penetrating the slope at an elevation of 80 meters; geometrically, the system calculated that the average slope of the upper part of the slope (60-100m) was 52°, which was marked as a high-risk steep slope area; hydrologically, the system found that the gullies in the upper part of the slope are the main water channels, which will directly scour the exposed area of ​​the weak interlayer; historically, the landslide repair area in 2019 was prominently marked, and its location is exactly at the intersection of the high-risk steep slope area and the weak interlayer.

[0025] Multidimensional analysis was initiated on the 3D model of the high slope: Geologically, the system automatically highlighted the weak interlayer penetrating the slope at an elevation of 80 meters; Geometrically, the system calculated that the average slope of the upper part of the slope (60-100m) was 52°, which was marked as a high-risk steep slope area; Hydrologically, the system found that the gullies in the upper part of the slope are the main water channels, which will directly scour the exposed area of ​​the weak interlayer; In terms of historical risk, the landslide repair area in 2019 was marked prominently, and its location is exactly at the intersection of the high-risk steep slope area and the weak interlayer.

[0026] When constructing the hierarchical system, the system discovered that PTZ camera A, deployed on the opposite mountainside, could perfectly cover SubZone-004 (high-risk area) in terms of position and focal length. Therefore, SubZone-004 was assigned to the first layer and assigned to camera A for 24-hour uninterrupted monitoring. SubZone-009 (low-risk area at the foot of the slope) and a small blind spot behind SubZone-004 (also marked as high-risk) were assigned to the second layer. When planning the drone flight path, the system ensured that the flight path could fly over these subspaces with the optimal overlap rate to acquire high-quality images for periodic comparison.

[0027] The final output list fragments from the system are as follows: Item 1: {ID:B-Slope-SubZone-004, Level: First Level, Boundary Coordinates: [(X1,Y1),(X2,Y2),…], Responsible Device:PTZ-Camera-A, Monitoring Frequency: 10 minutes / time}; Item 2: {ID:B-Slope-SubZone-009, Level: Second Level, Boundary Coordinates: [(X3,Y3),(X4,Y4),…], Responsible Device:UAV-Route-01, Monitoring Frequency: 1 time / week}.

[0028] refer to Figure 3 In step S12, the specific steps are as follows: S121: In the multi-layer sub-monitoring space, the matching coefficient of each sub-monitoring space is determined according to the spatial location of each sub-monitoring space, the ground camera and the drone. Based on the matching coefficient of each sub-monitoring space, the corresponding spatial area, the ground camera and the drone, a multi-dimensional monitoring system for the high slope is constructed. At this time, the ground camera and the drone are respectively responsible for monitoring the corresponding sub-monitoring space. S122: Real-time monitoring of a multi-dimensional monitoring system. At this time, the first detection system is determined based on the detection range of the ground camera, the corresponding movement trajectory, and the corresponding sub-monitoring space, and the ground camera is triggered to perform the first visual detection on the sub-monitoring space. At the same time, the second detection system is determined based on the detection range of the drone, the corresponding flight trajectory, and the corresponding sub-monitoring space, and the drone is triggered to perform the second visual detection on the sub-monitoring space.

[0029] In the embodiments of this application, in a multi-layered sub-monitoring space, the matching coefficient of each sub-monitoring space is determined based on its spatial location, ground camera, and drone. A multi-dimensional monitoring system for high slopes is constructed based on the matching coefficient of each sub-monitoring space, its corresponding spatial area, ground camera, and drone. In this case, the ground camera and drone are responsible for monitoring their respective sub-monitoring spaces, taking into account the spatial location of each sub-monitoring space, the ground camera, and the drone as a whole, thus ensuring the accuracy of the matching coefficient of each sub-monitoring space.

[0030] At this point, the matching coefficient is a comprehensive score (normalized to between 0 and 1) used to quantify the suitability of a specific monitoring platform for a specific sub-monitoring space. It is a multi-factor weighted model, mainly including: geometric coverage factor (F_geo), which assesses whether the platform can completely cover the sub-space without obstruction; resolution factor (F_res), which assesses the ground image resolution that the platform can achieve; timeliness factor (F_time), which assesses the frequency and real-time nature of the data provided by the platform; environmental adaptability factor (F_env), which assesses the platform's ability to work under adverse weather and lighting conditions; and cost-effectiveness factor (F_cost), which assesses the resource consumption of performing a single task. The specific score is calculated through a weighted formula (e.g., matching coefficient = W1 × F_geo + W2 × F_res + ...), where the weights can be adjusted according to the specific requirements of the monitoring task (e.g., high-risk areas place greater emphasis on resolution and timeliness).

[0031] The system prioritizes subspaces based on their risk level and area, typically assigning the highest priority to high-risk, small-area areas. It then iterates through the sorted list of subspaces, selecting the monitoring platform with the highest matching coefficient for each subspace. For example, high-risk, small-area areas are prioritized for ground cameras with high resolution and timeliness scores, while low-risk, large-area areas are assigned to drones with high geometric coverage and cost-effectiveness scores. For high-risk areas not covered by cameras, drones, despite their poor timeliness, are selected due to their irreplaceable coverage capabilities. During allocation, the system considers load balancing and redundancy to avoid single-point overload and assigns dual primary responsibility for the highest-risk areas to achieve cross-validation. Finally, it outputs a multi-dimensional monitoring system, specifying the responsible platform, monitoring frequency, and data quality requirements for each subspace.

[0032] Specifically, the system calculated matching coefficients for two typical sub-spaces of a high slope. SubZone-004 is a high-risk, small-area region (15m x 15m) located at an elevation of 80 meters, requiring monitoring of micro-cracks. SubZone-009 is a low-risk, large-area region (80m x 50m) located at the toe of the slope, primarily monitoring surface erosion. The calculation results show that for SubZone-004, ground camera A, due to its ability to achieve high resolution (2mm / pixel) and real-time monitoring via PTZ zoom, has a comprehensive matching coefficient (after adjusting the weights for high resolution and timeliness) as high as 0.93. While the drone has a higher resolution, its poor timeliness and weak environmental adaptability result in a matching coefficient of only 0.62. For SubZone-009, ground camera A, due to its long distance and low resolution, has a matching coefficient of only 0.35. The drone, however, can easily cover a large area, achieving a matching coefficient of 0.65.

[0033] Based on the matching coefficients mentioned above, the system begins to construct a monitoring framework. SubZone-004, with the highest priority, is processed. After comparing the matching coefficients, camera A (0.93) has a much higher coefficient than the drone (0.62). Therefore, SubZone-004 is assigned to ground camera A as the responsible platform, with a monitoring frequency of once every 10 minutes. Next, SubZone-009 is processed. The drone (0.65) has a much higher matching coefficient than camera A (0.35), so SubZone-009 is assigned to the drone, with a monitoring frequency of once a week. Finally, it is assumed that there is another high-risk SubZone-006. Located in the blind spot of camera A, its camera matching coefficient is 0, and the drone matching coefficient is 0.60. The system will assign it to a drone and increase the monitoring frequency to once every two days due to its high risk. The final multi-dimensional monitoring system configuration table clarifies the division of responsibilities for each subspace. For example, B-Slope-SubZone-004 (high risk) is the responsibility of PTZ-Camera-A, which monitors once every 10 minutes, and the data requirement is GSD≤3mm; B-Slope-SubZone-009 (low risk) is the responsibility of UAV-01, which monitors once a week, and the data requirement is GSD≤5cm.

[0034] Furthermore, a real-time multi-dimensional monitoring system is implemented. At this point, the first layer of detection is determined based on the detection range, corresponding movement trajectory, and corresponding sub-monitoring space of the ground camera, and the ground camera is triggered to perform the first layer of visual detection on the sub-monitoring space. Simultaneously, the second layer of detection is determined based on the detection range, corresponding flight trajectory, and corresponding sub-monitoring space of the drone, and the drone is triggered to perform the second layer of visual detection on the sub-monitoring space. This system takes into account the overall consideration of the drone's detection range, corresponding flight trajectory, and corresponding sub-monitoring space, ensuring the accuracy of the second layer of detection system.

[0035] At this point, for the first-level detection system, its detection range is determined, which is the set of all sub-monitoring spaces assigned to the camera in S121; its movement trajectory is determined, which is a preset, cyclically executed list of pre-set positions for inspection; the system calculates one or more optimal observation positions (i.e., preset positions, including pan, tilt, and zoom) for each assigned sub-space, and arranges all preset positions into an ordered list according to strategies such as risk priority or spatial proximity; the camera hardware, the set of sub-spaces it is responsible for, and the camera's dedicated list of pre-set positions together constitute a complete first-level detection system.

[0036] The first layer of detection is triggered by a high-precision task scheduler. The triggering is periodic, and its period is determined by the monitoring frequency defined in S121. The trigger is usually a timer based on the system clock, which performs a simple self-check when triggered. The execution process is as follows: the scheduler issues an instruction, and the camera starts from the standby position and moves in the order of the movement trajectory list. After reaching each preset position, it pauses, stabilizes, automatically focuses, and takes a high-resolution image with a timestamp and preset position label. After completing the shooting of all preset positions, the camera returns to the standby position and waits for the next trigger.

[0037] For the second layer of detection system, its detection range is determined, which is the geographical set of all sub-monitoring spaces assigned to the UAV in S121; its flight trajectory is determined, which is a three-dimensional route consisting of multiple waypoints; the system automatically generates an optimal route based on the boundaries of all assigned sub-spaces to ensure that it flies over all target areas with sufficient overlap, and sets the flight altitude, speed, camera attitude and motion parameters for each waypoint; the second layer of detection system is composed of the UAV hardware, the geographical set of the sub-spaces it is responsible for, and the UAV's exclusive preset route.

[0038] The second detection system is controlled by a multi-condition logic AND gate trigger. The task will only be triggered if all conditions are met simultaneously. These conditions include: time conditions (such as 10:00 AM every day), environmental conditions (wind speed, rainfall, and visibility queried via API), equipment status conditions (battery level, GPS signal), and airspace conditions (permit validity). The execution process is as follows: after all conditions are met, the system issues a command, the drone automatically takes off and flies autonomously along a preset route, automatically takes pictures at each waypoint, and transmits telemetry data back in real time; after the task is completed, it automatically returns to home and lands, and uploads all image data to the server.

[0039] Specifically, in the high slope case, SubZone-004 (high-risk remediation area) was assigned to PTZ camera A; the system calculated the optimal observation position for it (pan-tilt coordinates P=120°, T=25°, Z=20x) and saved it as Preset-004; at the same time, the camera is also responsible for monitoring a drainage ditch area (SubZone-011), for which the system generated Preset-007; the first detection system with PTZ camera A as the core was established, and its movement trajectory (inspection list) is: [standby position, Preset-004, Preset-007, standby position].

[0040] In the high slope case, SubZone-004 (high-risk remediation area) was assigned to PTZ camera A; the system calculated the optimal observation position for it (pan-tilt coordinates P=120°, T=25°, Z=20x) and saved it as Preset-004; at the same time, the camera was also responsible for monitoring a drainage ditch area (SubZone-011), for which the system generated Preset-007; the first detection system with PTZ camera A as the core was established, and its movement trajectory (inspection list) was: [standby position, Preset-004, Preset-007, standby position].

[0041] In the high slope case, SubZone-009 (a large area at the foot of the slope) and several other low-risk areas were assigned to the drone. The system planned an autonomous flight path with 10 waypoints for it, covering all assigned areas. The flight altitude was set at 120 meters to ensure a ground resolution better than 3cm. Thus, the second detection system with the drone as its core was established. Its responsibility is to periodically conduct wide-area aerial photography of all assigned areas, and the action plan is this preset 10 waypoint flight path.

[0042] The system began evaluating trigger conditions at 9:30:00 AM; the time conditions were met, the meteorological API showed a wind speed of 2 m / s and no rain, the drone status showed 95% battery and 12 GPS satellites, and the airspace permission was valid; at 10:00:00 AM, all conditions were met, and the system issued a takeoff command; the drone automatically took off and strictly followed the preset route to conduct a comprehensive second visual inspection of areas such as SubZone-009; after completing the mission, it automatically returned and landed at 10:20:00 AM and began uploading approximately 3GB of image data; through S122, the two detection systems worked together to form a dynamic monitoring network for high slopes around the clock with no blind spots.

[0043] refer to Figure 4 In step S13, the specific steps are as follows: S131: Real-time control of the first and second levels of monitoring, and output of multiple images. These images are transmitted along the image transmission channel to the image processing center. Based on the filtering of these multiple images, multiple key images of different dimensions are determined. Based on the image content of these key images, the corresponding sub-monitoring space, and the overall morphology of the high slope, the current anomaly area is predicted. S132: Based on the detection of the current abnormal area, determine the corresponding ground abnormal area and high-altitude abnormal area; trigger a second shot from the ground camera based on the ground abnormal area and a second shot from the drone based on the high-altitude abnormal area to output multiple second-shot images; determine a local image of the current abnormal area based on the synthesis of multiple second-shot images; determine the corresponding abnormal event based on the recognition of the local image.

[0044] In the embodiments of this application, the first and second levels of monitoring are controlled in real time, and multiple images are output. The multiple images are transmitted to the image processing center along the image transmission channel. Multiple key images of different dimensions are determined based on the screening of multiple images. The current abnormal area is predicted based on the image content of multiple key images, the corresponding sub-monitoring space and the overall shape of the high slope. This approach takes into account the overall consideration of screening multiple images and ensures the accuracy of multiple key images of different dimensions.

[0045] At this time, the system background service continuously monitors the operation status of the two detection systems, maintains a dynamic task queue, and records the start and end times of tasks, device status, and data volume to ensure the continuity of the data flow. Different transmission strategies are adopted for different data sources: high-frequency images output by the first detection system (camera) are transmitted in real time through the lightweight MQTT protocol, along with key metadata such as device ID, timestamp, and gimbal coordinates; periodic large-batch images output by the second detection system (drone) are uploaded in batches through the reliable HTTP / HTTPS protocol, and their metadata is stored in the accompanying file, including waypoint ID, GPS coordinates, etc.

[0046] The raw image stream contains a large amount of redundant and low-quality data. The purpose of filtering is to improve the efficiency and accuracy of subsequent processing. The filtering process covers multiple dimensions: Quality Analytical Analysis (IQA) eliminates invalid images that are blurry, overexposed, or severely occluded by calculating sharpness, exposure, and occlusion; Temporal Analytical Analysis maintains a historical baseline image library for each subspace and matches the best historical reference for new images; Spatial Analytical Analysis compares image metadata with subspace boundaries and retains only images that fully cover the target area. The key images obtained after the above multi-dimensional filtering are high-quality, comparable datasets that are precisely bound to specific subspaces, and serve as reliable input for subsequent intelligent analysis.

[0047] Through image content analysis, a Siamese network based on deep learning is used for change detection, outputting a change probability map. Simultaneously, a semantic segmentation model is used to identify specific land cover categories such as cracks and seepage. From a macroscopic perspective, subspace and overall morphological constraints are applied. Spatial labels of subspaces (such as risk level and main monitoring targets) are retrieved as prior knowledge weights, and two-dimensional anomaly regions are back-projected onto a three-dimensional model to analyze their spatial topological relationship with slope and geological structure. Combining microscopic and macroscopic analysis, the system finally outputs one or more structured current anomaly regions, which include boundary coordinates, anomaly type, confidence level, and associated subspace ID.

[0048] Specifically, at 10:00 AM, PTZ camera A completed an inspection, generating two images. These two images were instantly sent to the camera data queue in the image processing center via MQTT messages through the 5G network. The metadata indicated that one of the images was taken at P=120°, T=25°, Z=20x. At 10:20 AM, the drone completed an inspection, generating 300 aerial images, which were packaged into a ZIP file and uploaded to the drone data storage area in the processing center via HTTPS protocol, triggering a background processing task.

[0049] The image processing center received images from camera A. The system detected that one of the images had a high clarity score, normal exposure, and completely covered SubZone-004, and was therefore identified as a key image. Of the 300 images uploaded by the drone, 5 were removed during the quality screening stage due to uneven lighting caused by flying over clouds. The remaining 295 images underwent georegistration, and the system cropped out all image segments (15 in total) that covered SubZone-004. These segments were also identified as key images and were paired with the historical baseline images of SubZone-004.

[0050] The system inputs the current key image of SubZone-004 and the baseline image from last week into the Siamese network model. The model outputs a probability map of change, highlighting a thin, linear region. The semantic segmentation model confirms that the pixel category of this region is a crack. The system retrieves the attributes of SubZone-004 ({Risk Level: High, Main Target: Crack Expansion}) and projects this anomalous region onto the 3D model. It finds that it is located exactly at the edge of the 2019 landslide repair area and is almost parallel to the exposure line of the weak interlayer. After comprehensive analysis, the system predicts a current anomalous region: {ID:Anomaly-001, Boundary:[(x1,y1),…], Anomaly Type: New Crack, Confidence: 92%, Correlated Subspace:SubZone-004}. This structured Anomaly-001 object will be submitted for the next step of precise review and event confirmation.

[0051] Furthermore, based on the detection of the current anomaly area, corresponding ground anomaly areas and high-altitude anomaly areas are determined. Ground anomaly areas trigger secondary imaging by ground cameras, and high-altitude anomaly areas trigger secondary imaging by drones, outputting multiple secondary images. A local image of the current anomaly area is determined by synthesizing these multiple secondary images. The corresponding anomaly event is determined based on the recognition of this local image. This approach incorporates the overall consideration of synthesizing multiple secondary images, ensuring the accuracy of the local image of the current anomaly area. Simultaneously, a multi-dimensional monitoring system for high slopes is introduced to further control the first and second visual detection layers, incorporating the local image of the current anomaly area and improving the initial accuracy of anomaly events.

[0052] At this point, the system will determine whether the abnormal area is a ground-based or high-altitude abnormal area based on criteria such as the distance between the abnormal area and the camera, whether there is any obstruction, and whether the target being monitored needs 3D information. For ground-based abnormal areas, the system sends a high-priority pointing and shooting command to the corresponding PTZ camera, which automatically calculates the gimbal parameters and performs precise zoom shooting. For high-altitude abnormal areas, the system generates a detailed verification task and schedules the drone to perform multi-angle, high-resolution aerial photography tasks when safety conditions are met.

[0053] Images captured a second time need to be synthesized into a more informative local image product. For single close-up images (such as those taken by ground cameras), synthesis mainly refers to image optimization using methods such as super-resolution, noise reduction, and contrast enhancement to obtain the best visual effect and the clearest details. For multi-angle images (such as those taken by drones), photogrammetry is used to generate a high-density 3D point cloud model or digital surface model (DSM) of the anomalous area through structure of motion reconstruction (SfM) and multi-view stereo matching. This 3D model is the most informative local image.

[0054] For 2D enhanced images, the system applies subpixel-level precision image processing methods, such as edge detection and skeletonization, to accurately measure the length and width of cracks. For 3D point cloud models, it performs 3D point cloud analysis, estimating the depth of cracks or the volume of collapse bodies through plane segmentation and distance calculation. The system matches the identified quantified features with a pre-set anomaly event database to generate an anomaly event. The anomaly event includes event ID, type, location, geometric parameters, risk level, and discovery time, serving as the direct basis for submission to the next step of risk assessment.

[0055] Specifically, S131 predicts Anomaly-001 (newly formed crack); the system queries the device database and finds that the coordinates of the crack are 280 meters away from PTZ camera A in a straight line, with no obstructions in between; therefore, the system classifies it as a ground anomaly area and immediately sends a command containing precise coordinates to camera A; camera A completes the calculation within 0.5 seconds, the pan-tilt unit quickly rotates to the target location, and adjusts the optical zoom to 36x, capturing a high-resolution close-up image of the crack.

[0056] Assuming that Anomaly-001 also needs to assess its depth, the system also triggers the UAV to perform a detailed verification. The UAV hovers around the area and takes 20 photos from different angles. Using these 20 photos, the system generates a high-density 3D point cloud model of the crack area through SfM. This local image not only contains the two-dimensional shape of the crack, but also accurately reproduces the undulation of the surrounding rock mass and the depth of the crack.

[0057] The system analyzes the close-up image of the crack captured by camera A. Using subpixel-level edge detection and curve fitting, it accurately measures the crack length to be 15.3 cm and the maximum width to be 1.2 mm. The system identifies an anomaly and generates the following structured report: {Event ID: Event-001, Event Type: New Tension Crack, Location: {Longitude:…, Latitude:…, Elevation: 82.5 m}, Associated Subspace: SubZone-004, Geometric Parameters: {Length: 15.3 cm, Maximum Width: 1.2 mm, Orientation: NE45°}, Risk Level: Level III - Warning, Discovery Time: 2023-10-27 10:05:00, Associated Image: [CamA_Zoom_…jpg], Analyst: AI-Vision-System-v2.1}. The detailed Event-001 object will be submitted to S14 for further quantitative evaluation and grading.

[0058] refer to Figure 5 In step S14, the specific steps are as follows: S141: Dynamically detect abnormal events and identify multiple sub-abnormal contents during the detection process. Based on each sub-abnormal content, the corresponding abnormal location, and the corresponding local image, determine the corresponding abnormal features to mark multiple abnormal features. S142: Based on the identification of each abnormal feature, determine the feature location and corresponding feature shape of the abnormal feature, perform weighted processing on the feature location, corresponding feature shape and previous abnormal events of multiple abnormal features, and mark the abnormal coefficient of the abnormal event in the weighted processing; S143: Determine the first anomaly coefficient based on the anomaly coefficient of the abnormal event and the corresponding sub-monitoring space; determine the second anomaly coefficient based on the anomaly coefficient of the abnormal event, the working history of the ground camera, and the working history of the drone; and determine the anomaly level of the abnormal event based on the mapping table of the first anomaly coefficient, the second anomaly coefficient, and the anomaly level.

[0059] In the embodiments of this application, abnormal events are dynamically detected, and multiple sub-abnormal contents are identified during the detection process. Based on each sub-abnormal content, the corresponding abnormal location, and the corresponding local image, the corresponding abnormal features are determined to mark multiple abnormal features. This approach takes into account the overall consideration of each sub-abnormal content, the corresponding abnormal location, and the corresponding local image, ensuring the accuracy of the corresponding abnormal features.

[0060] At this point, in-depth intelligent analysis is performed on the local image output by S13 to identify all the independent physical phenomena constituting the anomaly. The dynamics are reflected in two aspects: first, the system simultaneously starts multiple professional models such as crack detection and seepage identification for parallel analysis; second, if historical data allows, the currently identified phenomena are compared with the previous detection results to determine whether they are new, expanding, or stable. Through instance segmentation and cluster analysis, the system distinguishes different physical phenomena and forms independent sub-anomalies, such as deconstructing a complex local instability event into a main landslide, trailing edge tension cracks, and slope toe bulging.

[0061] For different types of sub-anomalies, the system calls different feature extraction method libraries: for cracks, sub-pixel level edge detection and skeletonization are used to measure their length, width, direction and other geometric and morphological features; for collapsed bodies, their volume and area are estimated by calculating the volume difference between the 3D point cloud model and the original terrain model; for seepage areas, their spectral features and area changes are analyzed. Each extracted feature is transformed into a standardized, structured data object and assigned a unique ID. Its structure usually includes information such as feature ID, type, value, unit, and parent sub-anomaly content ID, laying the foundation for subsequent automated weighted calculations.

[0062] Specifically, a complex Event-002 (local instability) was discovered in SubZone-007 of the high slope; the system dynamically detected the local image of Event-002 (a 3D point cloud model generated by a UAV); the instance segmentation model identified and segmented an irregular collapse area on the rendered point cloud; the clustering method identified a long, continuous point cloud depression band at the rear edge of the collapse area; after comprehensive analysis, the system determined that Event-002 contained the following three sub-anomalies: Content-A (main landslide), Content-B (rear edge tension crack), and Content-C (slope toe bulging).

[0063] The system extracts and labels features from the three sub-anomalies of Event-002. For Content-A (the main landslide), the system uses 3D point computing to calculate its volume and area, labeling Feature-001 (type: volume, value: 12.5, unit: m³) and Feature-002 (type: projected area, value: 28.3, unit: m²). For Content-B (the trailing edge tension crack), the system accurately measures it on the 3D model, labeling Feature-003 (type: length, value: 8.7, unit: m). Feature-004 (Type: Maximum Visible Width, Value: 15, Unit: cm) and Feature-005 (Type: Direction, Value: NW320°, Unit: degrees); for Content-C (slope toe bulge), the system marks Feature-006 (Type: Maximum Bulge Height, Value: 0.4, Unit: m) by comparing with the original terrain; at this point, a macroscopic Event-002 has been completely deconstructed into 6 precise, quantified anomaly features. This structured data will be fully passed to the next step for risk-weighted calculation.

[0064] Furthermore, based on the identification of each abnormal feature, the feature location and corresponding feature shape of the abnormal feature are determined. The feature locations, corresponding feature shapes, and past abnormal events of multiple abnormal features are weighted and processed. In the weighting process, the abnormal coefficient of the abnormal event is marked. This takes into account the overall consideration of the identification of each abnormal feature and ensures the accuracy of the feature location, corresponding feature shape, and weighting of multiple abnormal features and past abnormal events.

[0065] At this point, in determining the location of features, the system back-projects the local coordinates of each feature into the global geodetic coordinate system and associates them with the spatial context in the S111 digital base to determine whether it is located in a key geological structure, engineering structure, or high-risk subspace. In determining the morphology of features, the system further interprets the feature values, classifying cracks as tension or shear, and collapses as rotational or planar collapses. It also uses the built-in morphology-risk mapping knowledge base to associate the identified morphology with the risk level.

[0066] A multi-factor weighted evaluation model integrates isolated feature values ​​and contextual information into a comprehensive risk score. This model can typically be expressed as: Anomaly Coefficient = F_pos × W_pos + F_morph × W_morph + F_hist × W_hist; where the location factor (F_pos) is quantified based on whether the feature is located in a critical position; the morphological factor (F_morph) is scored based on its morphological type and geometric size; the historical factor (F_hist) compares the current event with a historical event database through pattern matching, and scores it based on similarity and the final consequences of the historical events; the weights (W) can be dynamically adjusted according to geological conditions, climate, season, etc.

[0067] The system writes the final anomaly coefficient calculated in the previous step into the data structure of the anomaly event as a key attribute, while recording the calculation time and the main contributing risk factors; the output is an updated anomaly event object, which contains all the original information and this crucial quantitative indicator, preparing for the next step of level assessment.

[0068] Specifically, the system processes the anomalous features of Event-002 (local instability); for Feature-003 (tail edge tension crack), the system projects its location onto the 3D model and finds that it is exactly located on the weak interlayer exposure line identified in S111, and is located within the high-risk subspace SubZone-007; in terms of morphology, the system determines that it is a typical tension crack, with its direction parallel to the slope direction; for Feature-001 (main landslide), the system determines that it is a small-scale rotational landslide.

[0069] The system performs a weighted calculation on Event-002: the location factor (F_pos) scores 0.55 because it is located on a weak interlayer (+0.3) and in a high-risk subspace (+0.2); the morphology factor (F_morph) scores 1.5 based on the morphology and size of the tensile cracks and the volume of the rotating landslide; the history factor (F_hist) scores as high as 0.95 because a highly similar (95%) historical event was found in the historical database that ultimately worsened; assuming the weights are W_pos=0.3, W_morph=0.4, W_hist=0.3, then the anomaly coefficient C=0.3X0.55+0.4X1.5+0.3X0.95=1.05.

[0070] After the system completes the calculation, it marks the anomaly coefficient of Event-002 as 1.05. This coefficient, which is greater than 1, is a strong warning signal. The updated Event-002 object now contains this key attribute: {Event ID:Event-002,…,Anomaly coefficient:1.05,Calculation time:2023-10-2711:30:00,Main risk factors:[morphological factor,historical factor]}. This Event-002 object with the anomaly coefficient will be completely passed to S143 for final rating.

[0071] Therefore, a first anomaly coefficient is determined based on the anomaly coefficient of the abnormal event and the corresponding sub-monitoring space. A second anomaly coefficient is determined based on the anomaly coefficient of the abnormal event, the working history of the ground camera, and the working history of the drone. The anomaly level of the abnormal event is determined based on the mapping table of the first anomaly coefficient, the second anomaly coefficient, and the anomaly level. This approach takes into account the overall consideration of the mapping table of the first anomaly coefficient, the second anomaly coefficient, and the anomaly level, ensuring the accuracy of the anomaly level of the abnormal event.

[0072] At this point, the system presets a spatial risk calibration factor (K_zone) for each sub-monitoring space. This factor is set based on the inherent geological risk and historical disaster frequency of the area. By calculating the first anomaly coefficient (C1=CXK_zone), the risk of an anomaly occurring in a high-risk area (K_zone>1.0) will be amplified, while the threat of an anomaly occurring in a low-risk area (K_zone<1.0) will be appropriately reduced, thus more realistically reflecting its comprehensive threat.

[0073] The system will evaluate the working status and health of all monitoring devices in real time, forming a device status calibration factor (K_device). This factor comprehensively considers the device's recent calibration and maintenance records, real-time working status (such as image clarity and GPS signal), and data transmission quality. By calculating the second anomaly coefficient (C2=C1XK_device), the credibility of an anomaly detected by a device in poor condition or under harsh conditions will be lowered accordingly, and vice versa, ensuring that the final decision is based on reliable data.

[0074] The system maintains an anomaly level mapping table, which defines different risk threshold ranges and maps them one-to-one with specific anomaly levels (such as a four-level early warning model), color codes, and response suggestions. The system compares the calculated final anomaly coefficient (C2) with the mapping table, automatically determines its anomaly level, and marks this level as the final attribute on the anomaly event, directly triggering the corresponding response process.

[0075] Specifically, for Event-002, the anomaly coefficient C calculated by S142 is 1.05; the system queries the sub-monitoring space where it occurred, which is SubZone-007, a high-risk area containing weak interlayers and historically unstable, and its spatial risk calibration factor is set to K_zone=1.1; therefore, the first anomaly coefficient C1=1.05X1.1=1.155, and the risk coefficient is further amplified.

[0076] The system traced the data source of Event-002, which was discovered by ground camera A and confirmed by the drone. Camera A was in excellent condition (K_device_cam=0.98), but the drone experienced slight crosswinds and GPS signal fluctuations during its mission (K_device_uav=0.92). The overall device status calibration factor (with higher weight for the drone) was calculated to be K_device≈0.944. Therefore, the second anomaly coefficient C2=1.155X0.944≈1.090. Considering the slight uncertainty in the data source, the final risk coefficient was slightly reduced, but it is still at a high level.

[0077] The system calculates the final anomaly coefficient C2 of Event-002 as 1.090. By comparing it with the preset four-level warning mapping table, 1.090 ≥ 0.9, which meets the judgment condition of Level IV (Danger). The system finally marks the anomaly level of Event-002 as Level IV - Danger. This level will be used as the highest priority alarm and will be immediately sent to all relevant personnel through multiple channels, and will automatically trigger the preset automated handling process, such as blocking the road below the danger zone. Thus, the complete closed loop from data perception to intelligent decision-making is completely formed.

[0078] At this time, the high slope abnormal event level assessment table is shown in Table 1: Table 1. Assessment Table of Abnormal Event Levels for High Slopes

[0079] refer to Figure 6 In step S15, the specific steps are as follows: S151: Real-time monitoring of various abnormal events, simultaneous control of multiple abnormal events, determination of the corresponding abnormal area based on the detection of the abnormal event, and determination of the first level of abnormal control content based on the abnormality level of multiple abnormal events and the overall shape of the high slope. S152: Determine the second level of anomaly control content based on the anomaly areas of multiple anomalies and the overall shape of the high slope, and construct an anomaly control system for the high slope based on the first and second level of anomaly control content. S153: In the anomaly control system for high slopes, multiple anomaly nodes are identified based on the anomaly control system, and corresponding anomaly combinations are determined based on the tracing of each anomaly node. At this time, the unscheduled inspection path of the UAV is determined based on the anomaly combination, the corresponding anomaly area, and the UAV; the key detection area of ​​the ground camera is determined based on another anomaly combination, the corresponding anomaly area, and the ground camera.

[0080] In the embodiments of this application, each abnormal event is monitored in real time, and multiple abnormal events are controlled simultaneously. The corresponding abnormal area is determined based on the detection of the abnormal event. The first level of abnormal control content is determined according to the abnormality level of multiple abnormal events and the overall shape of the high slope. This approach takes into account the overall consideration of the abnormality level of multiple abnormal events and the overall shape of the high slope, ensuring the accuracy of the first level of abnormal control content.

[0081] At this time, the system maintains an anomaly situation board in the background with a 3D model as the base map, which overlays and displays anomalies in all activities in real time. The icon of each event intuitively reflects its level, location and evolution trend. The essence of synchronous control lies in a multi-event correlation analysis engine, which runs continuously and intelligently correlates all events through spatial correlation analysis (identifying neighboring events), temporal correlation analysis (analyzing the time sequence of occurrence), and causal chain inference (combined with hydrogeological models), thereby gaining insight into potential systemic risks.

[0082] The system utilizes the geographic coordinates stored in the event data to accurately locate anomalies in the 3D model through efficient spatial indexing. Based on the anomaly type (point, line, area), the system defines corresponding spatial ranges as anomaly areas, such as circular buffers for point events, buffer zones for linear anomalies, or actual polygonal boundaries for area anomalies. These areas are prominently displayed on the situation dashboard, allowing managers to intuitively see the distribution and impact range of risks in physical space.

[0083] The system incorporates a multi-layered decision-making logic engine, whose decision-making is based on the set of anomaly levels and the overall morphology of the high slope. The decision-making logic follows the highest priority rule: as long as any Level IV (hazard) event exists, the highest level emergency plan is directly triggered. If there is no Level IV event, the engine will analyze the combination of multiple Level III or Level II events and use a three-dimensional geological model to determine whether they constitute a potential sliding surface or threaten the same key object in space, thereby deciding whether to raise the overall risk level. The final output is a macro-strategy instruction set, which defines the overall response level, notification targets, and action plan for the current stage, rather than specific equipment actions.

[0084] Specifically, suppose there are three anomalous events on the high slope situation panel: Event-002 (Level IV, located in the upper high-risk area), Event-003 (Level III, located at the slope toe, manifested as abnormal seepage), and Event-004 (Level II, located in the middle low-risk area). The multi-event correlation analysis engine detects that Event-002 and Event-003 are spatially located in the same hydrogeological unit, and Event-003 occurs 30 minutes after Event-002. The system initially infers that Event-003 is the result of the development of Event-002 and marks them as a strongly correlated event group.

[0085] The system processes strongly correlated event groups; for Event-002 (landslide), the system defines an irregular polygonal anomaly region A with an area of ​​28.3 square meters based on its 3D point cloud boundary; for Event-003 (water seepage), the system defines an anomaly region B with an area of ​​5 square meters based on the image segmentation results; on the status board, these two regions are highlighted and connected by a dynamic dashed line to indicate their potential relationship to the management personnel.

[0086] The system immediately skips all complex combination analysis and directly triggers the highest-level decision-making logic; the generated macro-strategy instruction set (first-level anomaly control content) is as follows: Global strategy: Activate Level 1 emergency response; immediately send high-level alerts to all pre-set personnel (emergency command center, site manager, and head of G108 National Highway Management Office) via SMS, telephone, and APP; Recommended measures: immediately prepare to close the affected section of G108 National Highway, evacuate all irrelevant personnel in the work area below the slope, and notify the geological expert database to prepare to rush to the scene; this instruction set will serve as the highest directive to guide all subsequent specific actions.

[0087] Furthermore, based on the abnormal areas of multiple abnormal events and the overall shape of the high slope, the second level of anomaly control content is determined. Based on the first and second levels of anomaly control content, an anomaly control system for the high slope is constructed, which takes into account the overall consideration of the abnormal areas of multiple abnormal events and the overall shape of the high slope, ensuring the accuracy of the second level of anomaly control content.

[0088] At this point, the system uses a three-dimensional geological and terrain model to perform spatial topological relationship analysis on all abnormal areas, aggregating areas that are spatially adjacent, geologically related, or engineering-related into one or more key control units. For each unit, the system calls upon the built-in engineering response knowledge base to generate a series of specific, spatially located control measures. These measures are generally divided into three categories: engineering measures recommendations aimed at directly intervening in the danger, enhanced monitoring measures aimed at obtaining more accurate data, and resource allocation recommendations aimed at mobilizing human and material resources.

[0089] The system uses the first layer of content (macro strategy) of S151 as the top-level goal of the control system, and the second layer of content (specific measures) of S152 as the lower-level task list to achieve the goal. The final output of the anomaly control system is a structured data object that clearly defines the overall strategy, control units, specific action items, responsible parties for each action item, required resources, time limits and priorities. This system is dynamic and will be updated as the situation develops, ensuring that the intentions of the top level can be accurately and efficiently communicated and executed.

[0090] Specifically, the system takes over the analysis results from S151 and continues to process Event-002 (Level IV landslide) and Event-003 (Level III seepage). The system's three-dimensional topology analysis shows that the abnormal areas of these two events are located within the same potential sliding block cut by a weak interlayer, and therefore are aggregated into a key control unit-A. For this unit, the system generates a second layer of anomaly control content, including engineering measures (immediately waterproof the landslide area and dig emergency drainage ditches at the seepage points), monitoring measures (urgently install crack gauges at the cracks and implement intensive aerial monitoring every 30 minutes), and resource allocation recommendations (notify the patrol team and emergency warehouse to prepare corresponding materials).

[0091] The system integrates the macro strategy of S151 (initiating a Level 1 emergency response: preparing to close roads and evacuate personnel) with the specific measures of S152 to construct a structured anomaly control system. Its lower-level task list clearly lists: task: laying rainproof cloth, responsible person: patrol team No. 2, resources: 2 rolls of geotextile, time limit: 1 hour, priority: highest; task: installing crack gauges, responsible person: technicians of patrol team No. 2, resources: 3 sets of portable crack gauges, time limit: 2 hours, priority: high; task: intensive aerial photography monitoring, responsible person: UAV flight control center, resources: 1 spare UAV, time limit: immediate execution, priority: high.

[0092] Therefore, in the anomaly control system for high slopes, multiple anomaly nodes are identified based on the system's identification capabilities. Corresponding anomaly combinations are determined by tracing each node. Then, the unscheduled inspection path of the drone is determined based on this anomaly combination, the corresponding anomaly area, and the drone. Similarly, the key detection area of ​​the ground camera is determined based on another anomaly combination, the corresponding anomaly area, and the ground camera. This approach considers the overall characteristics of the other anomaly combination, the corresponding anomaly area, and the ground camera, ensuring the accuracy of the ground camera's key detection area. Furthermore, anomaly levels are introduced, enabling a holistic consideration of the anomaly levels of multiple anomalies, the corresponding anomaly areas, and the overall morphology of the high slope. This improves the accuracy of the high slope anomaly control system and triggers the drone's unscheduled inspection path and marks the key detection area of ​​the ground camera.

[0093] At this point, the system analyzes the control system and identifies each specific action item (such as laying rainproof cloth, installing crack gauges, and intensive aerial monitoring) as an independent abnormal node. Each node contains attributes such as task type, target area, required resources, and priority. The system traces back and performs logical clustering and grouping based on the node's attributes, especially the required resources and target area. The most important grouping criterion is resource type, forming drone task combinations, camera task combinations, etc., to prepare for subsequent equipment scheduling.

[0094] The system obtains the union of the abnormal regions corresponding to all nodes in the combination to form a total task area; it calls an advanced coverage path planning method to generate a highly customized flight path that can efficiently and completely cover the task area; unlike the fixed grid route of conventional inspection, this path is dynamically generated and will adaptively adjust according to the shape of the task area; the system packages the generated path (a series of waypoint coordinates, flight parameters, etc.) into a task file and sends it to the designated UAV so that it can take off and execute autonomously.

[0095] The system acquires the abnormal region corresponding to the node in the combination, which is usually a precise geometric shape; it sends a high-level instruction to the corresponding ground PTZ camera, the core of which is to define a key detection area polygon and specify a new working mode, such as area watch mode or area scan mode; in watch mode, the camera will continuously observe the area at high magnification; in scan mode, it will make small-amplitude periodic movements within the area. This instruction is dynamic and will be updated as the abnormal region changes, ensuring that the camera always focuses on the most critical part.

[0096] Specifically, the system analyzes the anomaly control system built by S152 and identifies multiple abnormal nodes, such as Node-001 (monitoring task, requires drone), Node-002 (monitoring task, requires camera), and Node-003 (engineering task, requires manual intervention). The system groups them according to resource type to obtain a drone anomaly combination (including Node-001) and a camera anomaly combination (including Node-002), preparing for subsequent scheduling.

[0097] The system handles abnormal combinations of UAVs and obtains the target of Node-001—focusing on the boundary polygon of control unit-A; the system calls the coverage path planning method to generate a zigzag detailed inspection route that closely surrounds the boundary of the unit, with a lower flight altitude and higher image resolution; the system immediately sends this irregular inspection route to the standby backup UAV, instructing it to take off immediately to perform an intensive detailed inspection task.

[0098] The system handles abnormal camera combinations and identifies the target of Node-002—a critical crack at the trailing edge of Event-002. Based on its coordinates and orientation, the system generates a rectangular area as a key detection zone. The system sends a command to PTZ camera A to switch its mode to area monitoring, specify the rectangular area, and set 30x zoom. Camera A immediately detaches from the normal inspection mode and enters continuous monitoring mode for the crack.

[0099] Please see Figure 7 , Figure 7This is a schematic diagram of the structural composition of a visual detection-based dynamic monitoring system for high slopes according to an embodiment of the present invention; the visual detection-based dynamic monitoring system for high slopes includes: The multi-layer sub-monitoring space module 21 is used to mark the current position of the high slope, determine the monitoring space of the high slope based on the current position of the high slope, the corresponding overall shape and previous monitoring events, and divide the high slope into multi-layer sub-monitoring spaces based on the visual detection of the monitoring space of the high slope. The visual inspection module 22 is used to match the multi-layer sub-monitoring space to the ground camera and the drone, and to build a multi-dimensional monitoring system for the high slope. In the multi-dimensional monitoring system, the ground camera performs the first visual inspection on the sub-monitoring space, and the drone performs the second visual inspection on the remaining sub-monitoring space. The abnormal event module 23 is used to determine multiple images based on the first and second monitoring, predict the current abnormal area based on the multiple images and the overall shape of the high slope, trigger secondary shooting by ground cameras and drones based on the current abnormal area, so as to determine the local image of the current abnormal area and determine the corresponding abnormal event. The anomaly level module 24 is used to determine multiple anomaly features based on the detection of the anomaly event, determine the anomaly coefficient of the anomaly event based on the feature location, corresponding feature shape and previous anomaly events of the multiple anomaly features, and mark the anomaly level of the anomaly event. The anomaly control system module 25 is used to construct an anomaly control system for high slopes based on the anomaly level of multiple anomalies, the corresponding anomaly areas, and the overall shape of the high slopes, in order to trigger the unscheduled inspection paths of drones and mark the key detection areas of ground cameras.

[0100] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A dynamic monitoring method for high slopes based on visual detection, characterized in that, include: Mark the current location of the high slope, determine the monitoring space of the high slope based on the current location of the high slope, the corresponding overall shape and previous monitoring events, and divide the high slope into multiple sub-monitoring spaces based on visual detection of the monitoring space; The multi-layered sub-monitoring spaces are matched with ground cameras and drones to construct a multi-dimensional monitoring system for high slopes. In the multi-dimensional monitoring system, ground cameras perform the first layer of visual inspection on the sub-monitoring spaces, and drones perform the second layer of visual inspection on the remaining sub-monitoring spaces. The multi-dimensional monitoring system specifies the responsible platform, monitoring frequency, and data quality requirements for each sub-space. Based on the first and second levels of monitoring, multiple images are determined. Based on these multiple images and the overall shape of the high slope, the current abnormal area is predicted. The current abnormal area is then used to trigger secondary shooting by ground cameras and drones to determine local images of the current abnormal area and identify the corresponding abnormal event. The abnormal event includes event ID, type, location, geometric parameters, risk level, and discovery time. Based on the detection of the abnormal event, multiple abnormal features are identified. Based on the feature location, corresponding feature shape and previous abnormal events of the multiple abnormal features, the abnormal coefficient of the abnormal event is determined, and the abnormal level of the abnormal event is marked. If there are multiple abnormal events, an abnormal control system for the high slope is constructed based on the abnormality level of the multiple abnormal events, the corresponding abnormal areas, and the overall shape of the high slope, so as to trigger the unscheduled inspection path of the drone and mark the key detection areas of the ground camera.

2. The dynamic monitoring method for high slopes based on visual detection according to claim 1, characterized in that, The current location of the marked high slope is determined based on its current location, corresponding overall shape, and past monitoring events. The monitoring space of the high slope is then divided into multiple sub-monitoring spaces based on visual detection of this monitoring space, including: Collect the location coordinates of the high slope and determine its current location. Based on the high slope's map information, determine the overall shape of the high slope. Based on the high slope's current location, corresponding overall shape, and previous monitoring events, determine the high slope's monitoring space. Visual inspection is performed on the monitoring space of high slopes, and the monitoring space of high slopes is dynamically monitored in different dimensions. The monitoring space of high slopes is spatially divided, and multiple spatial markers are determined in the process of division. Based on multiple spatial markers, the position of ground cameras and the flight path of drones, multiple sub-monitoring spaces are determined, and the monitoring range of each sub-monitoring space is marked.

3. The dynamic monitoring method for high slopes based on visual detection according to claim 1, characterized in that, The process involves matching multiple sub-monitoring spaces to ground cameras and drones, and constructing a multi-dimensional monitoring system for high slopes. Within this system, ground cameras perform a first layer of visual inspection of the sub-monitoring spaces, while drones perform a second layer of visual inspection of the remaining sub-monitoring spaces. This includes: In the multi-layer sub-monitoring space, the matching coefficient of each sub-monitoring space is determined according to the spatial location of each sub-monitoring space, ground cameras and drones. Based on the matching coefficient of each sub-monitoring space, the corresponding spatial area, ground cameras and drones, a multi-dimensional monitoring system for high slopes is constructed. At this time, the ground cameras and drones are responsible for monitoring the corresponding sub-monitoring spaces. The real-time monitoring multi-dimensional monitoring system determines the first layer of detection based on the detection range, corresponding movement trajectory, and corresponding sub-monitoring space of the ground camera, and triggers the ground camera to perform the first layer of visual detection on the sub-monitoring space; at the same time, the second layer of detection is determined based on the detection range, corresponding flight trajectory, and corresponding sub-monitoring space of the drone, and triggers the drone to perform the second layer of visual detection on the sub-monitoring space.

4. The dynamic monitoring method for high slopes based on visual detection according to claim 1, characterized in that, The process involves determining multiple images based on the first and second levels of monitoring, predicting the current anomaly area based on these images and the overall morphology of the high slope, and triggering secondary imaging by ground cameras and drones based on this anomaly area to determine local images of the current anomaly area and identify the corresponding anomaly events, including: The first and second levels of monitoring are controlled in real time, and multiple images are output. These images are transmitted to the image processing center along the image transmission channel. Based on the screening of these multiple images, multiple key images of different dimensions are determined. The current abnormal area is predicted based on the image content of these key images, the corresponding sub-monitoring space, and the overall shape of the high slope.

5. The dynamic monitoring method for high slopes based on visual detection according to claim 4, characterized in that, The process of determining multiple images based on the first and second levels of monitoring, predicting the current anomaly area based on the multiple images and the overall shape of the high slope, triggering secondary shooting by ground cameras and drones based on the current anomaly area to determine local images of the current anomaly area and identify the corresponding anomaly event, also includes: Based on the detection of the current abnormal area, the corresponding ground abnormal area and high-altitude abnormal area are determined. The ground abnormal area triggers a second shot by the ground camera, and the high-altitude abnormal area triggers a second shot by the drone to output multiple second-shot images. The local image of the current abnormal area is determined by synthesizing the multiple second-shot images. The corresponding abnormal event is determined based on the recognition of the local image.

6. The dynamic monitoring method for high slopes based on visual detection according to claim 1, characterized in that, The process of determining multiple abnormal features based on the detection of the abnormal event, determining the abnormality coefficient of the abnormal event based on the feature location, corresponding feature morphology, and previous abnormal events of the multiple abnormal features, and marking the abnormality level of the abnormal event includes: Dynamically detect abnormal events and identify multiple sub-abnormal contents during the detection process. Based on each sub-abnormal content, the corresponding abnormal location, and the corresponding local image, determine the corresponding abnormal features to mark multiple abnormal features. Based on the identification of each abnormal feature, the feature location and corresponding feature shape of the abnormal feature are determined. The feature locations, corresponding feature shapes of multiple abnormal features and previous abnormal events are weighted and processed, and the abnormal coefficient of the abnormal event is marked in the weighting process.

7. The dynamic monitoring method for high slopes based on visual detection according to claim 6, characterized in that, The process of determining multiple abnormal features based on the detection of the abnormal event, determining the abnormality coefficient of the abnormal event based on the feature location, corresponding feature morphology, and previous abnormal events of the multiple abnormal features, and marking the abnormality level of the abnormal event, further includes: The first anomaly coefficient is determined based on the anomaly coefficient of the abnormal event and the corresponding sub-monitoring space. The second anomaly coefficient is determined based on the anomaly coefficient of the abnormal event, the working history of the ground camera, and the working history of the drone. The anomaly level of the abnormal event is determined based on the mapping table of the first anomaly coefficient, the second anomaly coefficient, and the anomaly level.

8. The dynamic monitoring method for high slopes based on visual detection according to claim 1, characterized in that, If there are multiple abnormal events, an anomaly control system for the high slope is constructed based on the anomaly level of the multiple abnormal events, the corresponding abnormal areas, and the overall shape of the high slope. This system triggers unscheduled inspection paths for drones and marks key detection areas for ground cameras, including: Real-time monitoring of various abnormal events, simultaneous control of multiple abnormal events, determination of the corresponding abnormal area based on the detection of the abnormal event, and determination of the first level of abnormal control content based on the abnormality level of multiple abnormal events and the overall shape of the high slope. The second level of anomaly control content is determined based on the abnormal areas of multiple abnormal events and the overall shape of the high slope. An anomaly control system for the high slope is then constructed based on the first and second levels of anomaly control content.

9. The dynamic monitoring method for high slopes based on visual detection according to claim 8, characterized in that, If there are multiple abnormal events, an anomaly control system for the high slope is constructed based on the anomaly level of the multiple abnormal events, the corresponding abnormal areas, and the overall shape of the high slope. This system triggers unscheduled inspection paths for drones and marks key detection areas for ground cameras. It also includes: In the anomaly control system for high slopes, multiple anomaly nodes are identified based on the identification of the anomaly control system, and corresponding anomaly combinations are determined by tracing each anomaly node. At this time, the unscheduled inspection path of the drone is determined based on the anomaly combination, the corresponding anomaly area, and the drone; and the key detection area of ​​the ground camera is determined based on another anomaly combination, the corresponding anomaly area, and the ground camera.

10. A dynamic monitoring system for high slopes based on visual detection, characterized in that, The visual detection-based dynamic monitoring system for high slopes is applied to the visual detection-based dynamic monitoring method for high slopes as described in any one of claims 1-9.

Images