Airport perimeter inspection methods and perimeter inspection robots
By combining a perimeter inspection robot with multimodal sensors and a unified time reference, continuous inspection of the airport perimeter can be located, quantified, and traced, solving the problems of blind spots and false alarms in existing technologies and improving the accuracy and safety of the inspection.
Patent Information
- Application Number
- CN202611566025.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-10
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies are insufficient for conducting continuous, quantifiable, and traceable inspections of airport perimeters in complex terrain and electromagnetic environments. Furthermore, fixed cameras and mobile inspection equipment suffer from blind spots, numerous false alarms, and difficulties in quickly verifying alarm points.
By employing a perimeter inspection robot, based on a segmented map of the airport perimeter and a unified time reference, and through closed-loop processing of positioning, detection, identification, alarm, and recording, combined with multimodal sensors such as visible light, infrared, and lidar, the robot can accurately identify and classify the status of the perimeter fence and intrusion targets.
It improves the verifiability and accuracy of airport perimeter inspections, reduces the false alarm rate, and ensures reliable transmission of security control and alarm information in complex environments.
Smart Images

Figure CN123113899A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of airport perimeter security inspection technology, and more specifically, to an airport perimeter inspection method and a perimeter inspection robot. Background Technology
[0002] Airport perimeter fencing is continuously deployed along the airport's outer perimeter, covering a long distance and a wide area, traversing patrol roads, grasslands, muddy ground, drainage ditches, slopes, and uneven surfaces. Inspection work is also affected by environmental factors such as nighttime, rain, snow, fog, haze, strong winds, and low light conditions. Currently, airport perimeter fencing inspections mainly rely on manual observation of the mesh panels, posts, and connectors. When loosening or instability is suspected, the fencing is shaken by hand. This method is labor-intensive, has long inspection cycles, and relies on personnel experience, making it difficult to create structured records that correlate fencing segment numbers, coordinates, and detection waveforms. While fixed cameras, perimeter vibration sensors, and ordinary mobile inspection equipment can provide some images or alarm information, they suffer from problems such as blind spots, numerous false alarms, difficulty in quickly verifying alarm points, and inconsistencies in the spatiotemporal references of various sensors. Visual inspection alone is insufficient to confirm loose mesh panels, broken wires, detached welds, deformation, or post instability. Furthermore, contact force thresholds alone are insufficient to distinguish between localized mesh loosening and low-frequency swaying of the supporting structure.
[0003] Therefore, how to conduct continuous, locatable, quantifiable, and traceable inspections of airport perimeters in complex terrain and electromagnetic environments, and how to ensure the transmission of security control and alarm services when communication conditions change, are urgent problems that need to be solved. Summary of the Invention
[0004] In view of this, the purpose of this application is to provide an airport perimeter inspection method and a perimeter inspection robot, which can improve the problem that multiple detection modules operate independently and are difficult to verify each other.
[0005] In a first aspect, embodiments of this application provide an airport perimeter inspection method applied to a perimeter inspection robot, comprising: determining the perimeter inspection robot's segment number and positioning coordinates based on a pre-established airport perimeter segment map and a unified time reference; controlling the perimeter inspection robot to enter a safe approach state and adjust its body posture under safe conditions; synchronously collecting the current motion parameters of the perimeter inspection robot; determining the effective displacement of the perimeter fence along the loading direction based on the current motion parameters; determining the perimeter fence state characteristic quantity based on the effective displacement, and outputting the perimeter fence defect category based on the deviation between the perimeter fence state characteristic quantity and the normal perimeter fence response reference corresponding to the perimeter fence segment number; collecting image information of the perimeter fence's inner and outer areas, and processing the image information... The system performs target identification, determines the target category of the fence intrusion target, and establishes the target trajectory of the fence intrusion target. Based on the fence centerline and inner / outer directions of the fence segment associated with the fence intrusion target, a fence centerline determination surface is established. Based on the spatial relationship between the target trajectory and the fence centerline determination surface, it is determined whether the fence intrusion target is located outside the fence, in the fence neighborhood, inside the fence, or has effectively crossed from the outside to the inside, and personnel and biological intrusion identification results are output. According to the fence segment number and time window, the physical inspection results corresponding to the fence defect category are correlated with the personnel and biological intrusion identification results, and the alarm level is determined based on the confirmed intrusion, physical state, or security risk event. According to the alarm level, the alarm evidence package and inspection records are uploaded to the airport inspection management platform.
[0006] In the above implementation process, by organizing the positioning, approach, detection, diagnosis, identification, alarm and recording stages using the segmented map of the airport perimeter and a unified time reference, the output of the previous stage becomes the trigger condition or constraint condition of the next stage. The spatial relationship between the target and the perimeter is transformed into a determinable motion state by using the perimeter centerline determination surface. This can improve the problem that multiple detection modules operate independently and are difficult to verify with each other, and is conducive to forming a traceable perimeter inspection closed loop.
[0007] In one embodiment, establishing a fence centerline determination surface based on the fence centerline and inner / outer directions of the fence segment associated with the fence intrusion target includes: obtaining the fence centerline, inner / outer directions, and allowed association range of the fence segment associated with the fence intrusion target stored in the airport boundary segmentation map, and establishing the fence centerline determination surface based on the fence centerline and the inner / outer directions; the method further includes: forming an outer trajectory record when the fence intrusion target is located outside the fence and moves parallel to it; forming an approach trend record when the fence intrusion target continues to approach the fence; wherein, the fence centerline determination surface, inner / outer markings, and the target trajectory are used for boundary spatial relationship identification.
[0008] In the above implementation process, the fence centerline determination surface is established by storing the fence centerline, inner and outer directions and allowed association range in the airport perimeter segment map. The spatial state of the target is quantified into outer trajectory records, approach trend records and trajectory basis for climbing, crossing, gathering and intrusion identification. This avoids making intrusion conclusions directly based on single frame images, single thermal image contours, single vibrations or single point clouds, which helps to improve the verifiability of intrusion identification and reduce false alarms.
[0009] In one embodiment, the method further includes: establishing a continuous trajectory for airborne targets in the areas inside and outside the fence, and projecting the trajectory of the airborne targets onto the fence centerline determination surface, fence segment, and controlled area inside the fence; forming an airborne object intrusion identification result when the airborne target enters the fence neighborhood from the outside of the fence and continues to move towards the inside of the fence, or enters the controlled area and maintains a traceable trajectory; outputting an unidentified airborne target when the category confidence of the airborne target is insufficient; and classifying the airborne target as a foreign object suspension inspection when the airborne target is only suspended on the fence mesh, sign, or connecting component and does not form a continuous boundary crossing trajectory from the outside to the inside.
[0010] In the above implementation process, by establishing a continuous trajectory for the airborne target and projecting it onto the judgment surface, fence section, and controlled area, the airborne object intrusion identification result is output only when a traceable trajectory is formed from the outside to the inside or when it enters the controlled area. Airborne targets that are only suspended and have not crossed the boundary are classified as foreign object suspension inspections, avoiding direct false alarms of fence attachments as intrusions. At the same time, the multimodal combination of visible light, lidar, and infrared or thermal imaging improves the quality of airborne target category and spatial relationship identification.
[0011] In one embodiment, the method further includes: spatiotemporally aggregating the trajectories of multiple individuals located in the vicinity outside the fence to determine the number of individuals, spatial density, dwell time, consistency of movement direction, distance changes from the centerline of the fence, and approach trend of the multiple individuals; outputting a crowd attention event when the multiple individuals do not approach the fence, do not exhibit climbing characteristics, do not exhibit crossing characteristics, and do not form an effective association with abnormal vibrations in the same fence segment; and re-establishing an intrusion identification event centered on the trajectory of any one of the multiple individuals when any one of the individuals continuously approaches the fence, exhibits climbing characteristics, exhibits crossing characteristics, or forms an effective association with the abnormal vibrations.
[0012] In the above implementation process, by spatiotemporally aggregating the trajectories of multiple people in the outer neighborhood and distinguishing between crowd attention events and intrusion identification events, the security alarms for group gatherings are retained, while avoiding directly treating the gathering of people who have not approached, climbed, or crossed as illegal intrusions.
[0013] In one embodiment, the method further includes: when the outline of the fence intrusion target continuously overlaps with the fence area, the target trajectory continuously changes along the fence height direction and conforms to climbing characteristics, and the fence vibration sensor of the same fence segment or adjacent fence segments allowed by calibration outputs abnormalities within the associated time window, forming a climbing probability and identification quality mark of the fence intrusion target, and recording the trajectory identifier, fence segment number, route coordinates, shortest distance, fence overlap continuity, height direction change characteristics, movement direction characteristics, vibration correlation, and input quality of the fence intrusion target; when the fence vibration quality is unqualified, the fence intrusion target is severely obstructed, the inner and outer markings are unstable, or the trajectory continuity is insufficient, outputting "person climbing pending verification"; when the fence intrusion target is a person and is located near the fence, controlling the fence inspection robot to maintain a safe distance for observation, without approaching, loading, or contacting the fence; after the person leaves and the safety conditions are restored, arranging a physical state review of the fence.
[0014] In the above implementation process, the overlapping of contours, continuous changes in height direction and abnormal vibration of vibration sensors in the same or adjacent fence segments are used as climbing judgment conditions. When the vibration quality, obstruction or trajectory continuity is insufficient, the system outputs that the person is climbing and needs to be verified, thus avoiding the identification of climbing based on a single image clue. At the same time, the robot is prohibited from approaching, loading or touching the fence when the person is near it, and physical verification is arranged only after the person leaves and the safety conditions are restored, which helps to ensure personal safety.
[0015] In one embodiment, the method further includes: performing a continuity check on the target trajectory before, during, and after crossing to exclude apparent crossings caused by temporary obstruction, viewpoint switching, location jumps, or unstable fence segment affiliation; outputting a personnel intrusion identification result or a biological intrusion identification result when the target trajectory continuously enters the inner side of the fence from the outer side, the fence intrusion target remains within the fence boundary, and effectively crosses the fence centerline determination plane; and outputting a verification result when the fence intrusion target only briefly enters the fence neighborhood, does not effectively cross the fence centerline determination plane, cannot be continuously tracked, or the evidence quality is insufficient.
[0016] In the above implementation process, the continuity of the trajectory before, during and after crossing is checked to eliminate the apparent crossing caused by temporary occlusion, viewpoint switching, and positioning jump. The continuous entry into the inner side, the continuous location in the inner side and the effective crossing are used as combination conditions to improve the credibility of personnel and biological intrusion identification results and avoid identifying targets that enter the neighborhood for a short time or do not cross effectively as deterministic intrusions.
[0017] In one embodiment, the method further includes: verifying the consistency of the visible appearance information, trajectory, location, and current time of the fence intrusion target with the valid work permit information issued by the current task, and generating an authorized operator verification result; outputting "Authorized operator verification passed" when the fence intrusion target is located within the permitted work area and the valid time period, the visible appearance information is consistent with the parameter package, and the valid work permit information matches the current task; outputting "Authorized operator pending verification" when the valid work permit information is valid but the appearance features, trajectory continuity, visibility ratio, observation distance, or nighttime imaging quality are insufficient; and outputting "Personnel identity pending verification" when the fence intrusion target is located in a non-permitted area, the work time period is inconsistent, the permit status is invalid, or the visible evidence is insufficient, and continuing to apply personnel intrusion identification and event classification rules.
[0018] In the above implementation process, by verifying the consistency between the appearance of the work clothes, reflective vests, safety helmets, personnel trajectory, location, current time and valid work permit information, the system outputs "Authorized operator verified," "Authorized operator pending review," or "Personnel identity pending verification." It does not read personal identity information or perform facial recognition, which can meet the needs of operator management while avoiding personal identity verification.
[0019] In one embodiment, the method further includes: determining the current day / night mode based on environmental perception quality, a unified time reference, and a current parameter package; wherein, in daytime mode, a visible light camera is used as the primary identification source, with infrared or thermal imaging devices and lidar as auxiliary identification sources; in nighttime mode, passive vibrations from infrared or thermal imaging devices, lidar, and fence vibration sensors are used as the primary identification source, with visible light low-light images as auxiliary identification sources; for each identification event, the day / night mode, available modal set, image quality, thermal imaging validity, point cloud registration quality, positioning quality, time synchronization quality, vibration quality, trajectory continuity, fence segment binding quality, model version, and degradation reason are recorded; in nighttime mode, the event level that has met the same evidence conditions is not reduced due to modal switching; when the complete identification conditions cannot be met only because some modalities are unavailable, the identification results and quality labels of the currently available modalities are output; when the evidence is insufficient, the results to be verified are output.
[0020] In the above implementation process, by allocating primary and secondary identification sources according to day and night modes, and recording the available modal set and various quality markers for each identification event, results with quality markers or results to be verified can still be output at night when some modalities are unavailable or evidence is insufficient, avoiding the substitution of deterministic conclusions with a single clock condition or low-quality modalities. In one embodiment, determining the fence segment number where the fence inspection robot is located and the positioning coordinates of the fence inspection robot based on a pre-established airport perimeter segment map and a unified time reference includes: acquiring multiple positioning results and determining the credibility of each positioning result; wherein the credibility of the positioning results is configured to be updated based on cross-positioning source residuals; and dynamically weighting and fusing the positioning results that have passed the consistency check to output the current parameters of the fence inspection robot.
[0021] In the above implementation process, by maintaining the reliability of each positioning source with dynamic updates of the residuals across positioning sources, and by dynamically weighting and fusing the positioning results that have passed the consistency test, the problem of the single positioning method being degraded under conditions of occlusion, slippage, rain, fog or low light affecting the continuity of perimeter inspection can be improved, which is conducive to the stable control of the distance and attitude between the robot and the fence.
[0022] In one embodiment, the method further includes: when the confidence level drops to a first threshold, controlling the perimeter inspection robot to decelerate and increase the safe distance between it and the fence; when the confidence level drops to a second threshold, controlling the perimeter inspection robot to stop and reposition itself in place; wherein the second threshold is less than the first threshold; and in the event of repositioning failure, controlling the perimeter inspection robot to return along the confirmed trajectory or wait for remote takeover.
[0023] In the above implementation process, by slowing down according to the level of confidence, increasing the safety distance, stopping repositioning and returning or waiting for remote takeover, the risk of the robot continuing to approach the fence when the positioning confidence is insufficient can be reduced, which will lead to collisions or false detections and improve the safety of the fence inspection process.
[0024] In one embodiment, the step of performing target recognition on the image information, determining the target category of the fence intrusion target, and establishing the target trajectory of the fence intrusion target includes: acquiring image information of the area inside and outside the fence using a visible light camera, infrared or thermal imaging device; using a detection and tracking model trained on airport scene samples to identify at least one target among people, dogs, birds and other animals in the image information; establishing a continuous trajectory for each identified target and outputting target information; wherein, the target information includes: target category, detection and tracking confidence, position, speed, direction of movement, closest distance to the fence centerline, inner and outer markings, associated fence segment number, route coordinates, geographic coordinates, unified timestamp, and identification quality marker.
[0025] In the above implementation process, multimodal acquisition of the area inside and outside the fence is carried out by using visible light cameras and infrared or thermal imaging equipment. The detection and tracking model trained with airport scene samples is used to identify people, dogs, birds and other animals. A continuous trajectory is established for each target and the category, confidence level, position, speed, direction of movement, closest distance to the center line of the fence and the inner and outer marks are output. The image information is transformed into structured target information that can be associated with the fence segment number and timestamp, which is conducive to providing evidence of intrusion targets under the same spatiotemporal index for subsequent cross-modal fusion.
[0026] In one embodiment, the method further includes: determining the climbing probability and crossing probability of the target trajectory based on the relationship between the target trajectory and the centerline of the fence; wherein the detection result of the fence intrusion target includes at least one of the target category, the climbing probability, the crossing probability, and the vibration anomaly probability corresponding to the fence vibration signal.
[0027] In the above implementation process, by establishing the target trajectory and determining the climbing probability and crossing probability based on its relationship with the center line of the fence, the visual detection results can participate in subsequent fusion in the form of probabilistics, which can improve the problem that it is difficult to participate in quantitative fusion when only the target category is output.
[0028] In one embodiment, determining the climbing probability and crossing probability of the target trajectory based on the relationship between the target trajectory and the centerline of the fence includes: constructing a crossing determination surface using the centerline of the fence and the inner and outer directions; generating a attention event when the fence intrusion target is located outside the fence and moves parallel to it; increasing the attention level of the fence intrusion target when the fence intrusion target continues to approach the fence; forming the climbing probability when the outline of the fence intrusion target overlaps with the fence area and the height or position change of the target trajectory conforms to climbing characteristics; and forming the crossing probability when the target trajectory continuously enters the inner side of the fence from the outer side and satisfies spatial and duration conditions.
[0029] In the above implementation process, by constructing a crossing determination surface with the fence centerline and the inner and outer directions, the motion state of the target relative to the fence is quantified into the event of concern, the level of concern, the climbing probability, and the crossing probability, respectively. This makes the determination of intrusion behavior verifiable in space and time conditions, which is conducive to converting the intrusion detection results into probabilistic evidence that can participate in quantitative fusion before cross-modal fusion.
[0030] In one embodiment, controlling the perimeter inspection robot to enter a safe approach state and adjust its body posture when safety conditions are met includes: projecting real-time positioning coordinates onto the target fence segment to determine the lateral and heading errors of the perimeter inspection robot relative to the inspection offset line; determining the fence normal according to the fence direction associated with the fence segment number; and, when the safety conditions are met, deploying the side touch detection module and adjusting the body posture so that the side touch detection module contacts the fence along the fence normal; and, when the safety conditions are not met, controlling the perimeter inspection robot to reposition itself or collect preset data.
[0031] In the above implementation process, by projecting the real-time positioning coordinates onto the target fence segment and determining the contact direction with the fence normal, the touch loading direction is consistent with the force direction of the fence, which can improve the problem of incomparable detection waveforms caused by contact point offset or loading direction deviation; when the safety conditions are not met, it switches to passive acquisition, which is beneficial to retain usable evidence in fence segments where active touch is not possible.
[0032] In one embodiment, the current motion parameters include: contact force, displacement of the side touch detection module, and base displacement; determining the effective displacement of the fence along the loading direction based on the current motion parameters includes: compensating for the displacement of the side touch detection module and the base displacement based on the equivalent flexibility of the side touch detection module and the contact force to obtain the effective displacement; the equivalent flexibility is obtained by graded loading tests on a rigid calibration member and grouped according to temperature or mechanism state; wherein the base displacement is jointly estimated by the kinematics, inertial measurement attitude, and short-time positioning increment of the fence inspection robot.
[0033] In the above implementation process, by compensating for the elastic deformation of the touch mechanism and the motion of the robot base respectively, the displacement measured by the actuator is converted into the effective displacement of the fence, so that the force-displacement characteristics obtained by different actions and different devices have a comparable basis, which helps to reduce the influence of the robot's own motion mixed into the detection results.
[0034] In one embodiment, determining the fence state characteristic quantity based on the effective displacement includes: determining the equivalent stiffness and normalized hysteresis based on the contact force and the effective displacement, and determining the residual displacement after unloading; performing time-frequency decomposition on the fence vibration signal after removing the base movement to determine the energy ratio and vibration attenuation coefficient of each frequency band; and determining the fence state characteristic quantity based on at least one of the equivalent stiffness, the normalized hysteresis, the residual displacement, the energy ratio, the vibration attenuation coefficient, and the spatial consistency of adjacent detection points.
[0035] In the above implementation process, the fence state characteristic quantity is formed by combining the force-displacement curve characteristics, unloading residue, vibration time-frequency energy and attenuation characteristics, and spatial consistency of adjacent detection points. The empirical inspection of manually shaking the fence is transformed into repeatable and calibrable quantitative characteristics, which helps to improve the problem that a single contact force threshold is difficult to distinguish the defect type.
[0036] In one embodiment, the method further includes: determining a baseline group according to the fence parameters corresponding to the fence segment number; standardizing the fence state feature quantity using the median and median absolute deviation of the baseline group; returning to the previous level baseline group if the baseline samples are insufficient; and stopping the output of the fence defect category if the data quality is unqualified.
[0037] In the above implementation process, benchmarks are established by grouping according to structural type, material, column spacing, temperature range and wind conditions, and robust standardization is performed using median and median absolute deviation to reduce the interference of abnormal samples on the benchmarks. When the benchmark samples are insufficient, the data is returned to the previous level benchmark group and the diagnostic confidence is reduced and the benchmark level is recorded. When the data quality is unqualified, the output of the defect category is stopped, which helps to avoid outputting unfounded definitive conclusions.
[0038] In one embodiment, determining the fence state characteristic quantity based on the effective displacement and outputting the fence defect category based on the deviation between the fence state characteristic quantity and the normal fence response benchmark corresponding to the fence segment number includes: determining the fence defect category as mesh loosening when the equivalent stiffness is reduced relative to the normal fence response benchmark, and at least one of the following is true: the normalized hysteresis is increased, the vibration attenuation coefficient is decreased, or the energy ratio exceeds the normal frequency band, and the anomaly occurs at the mesh detection point; determining the fence defect category as wire breakage or weld failure when a sudden change occurs during loading and is consistent with the same positional change in the preset state; wherein the preset state includes: visual linear gap, missing connection point, or adjacent repeated action; determining the fence defect category as mesh deformation when the residual displacement exceeds the residual displacement threshold and the zero force position of the repeated touch action maintains the same directional deviation; determining the fence defect category as column instability when low-frequency displacement in the same direction occurs on both sides of the same column or at multiple adjacent detection points, the compensated base displacement is insufficient to explain the low-frequency displacement, and the low-frequency energy or attenuation characteristics exceed the normal benchmark of the column.
[0039] In the above implementation process, by setting executable criteria for mesh loosening, wire breakage or detachment, mesh deformation and column instability, and by using spatial gating of "local mesh anomaly" and "consistent anomaly across adjacent detection points" to distinguish between mesh loosening and column instability, the problem that local loosening and low-frequency oscillation of the support structure cannot be distinguished by contact force threshold alone can be improved, which is conducive to improving the distinguishability of defect category judgment.
[0040] In one embodiment, the step of associating the physical inspection results corresponding to the fence defect category with the personnel and biological intrusion identification results according to the fence segment number and time window, and determining the alarm level based on the confirmed intrusion, physical state, or security risk events, includes: determining the personnel and biological intrusion identification results, the physical inspection results corresponding to the fence defect category, and the fence vibration anomaly results with the same fence segment number or belonging to adjacent fence segments that are allowed by calibration, and whose time difference is within the time window, as an event association set; classifying the confirmed events into personnel or biological intrusion events, personnel climbing events, high-intensity impact events, fence physical state anomaly events, and passage risk events based on the event association set; and generating observation, attention, warning, or emergency alarms based on the confirmed event types.
[0041] In one embodiment, the method further includes: if any submodule of the personnel and biological intrusion identification, the fence defect category diagnosis, or the fence vibration detection is unavailable or has substandard identification quality, the identification result of the available modality or the available submodule is downgraded to participate in the event association, and the reason for the missing, the reason for the downgrade, and the quality mark are recorded; if the safety conditions do not allow active touch control, the alarm level is determined solely based on visual, thermal imaging, and passive vibration evidence.
[0042] In the above implementation process, by establishing event associations according to fence segment numbers and time windows, and by unifying the identification results of personnel and biological intrusions, the physical inspection results corresponding to fence defect categories, and the fence vibration anomaly results into an event association set, and outputting observation, attention, warning, or emergency alarms according to the confirmed intrusion, physical state, or security risk event types, the problem of inconsistent spatiotemporal references of data from different sensors and the difficulty in cross-verifying can be improved. By classifying by event type instead of relying on a single confidence threshold, it is beneficial to suppress false alarms caused by a single visual or single vibration noise without masking emergency events.
[0043] In one embodiment, the perimeter inspection robot is a multi-legged wheeled robot, which includes a body, multiple mechanical legs, and drive wheels disposed at the ends of each mechanical leg. The method further includes: establishing a local elevation map based on laser signals and visual signals, and determining a terrain score of the terrain where the multi-legged wheeled robot is located by combining the current motion parameters and terrain parameters of the multi-legged wheeled robot; determining the motion mode of the multi-legged wheeled robot based on the terrain score; wherein the motion mode includes: a high-speed wheeled driving mode, a low-speed wheeled driving mode, a wheel-leg cooperative obstacle-crossing mode, and a safe stopping mode.
[0044] In the above implementation process, by determining the terrain score based on the local elevation map and the current motion parameters and terrain parameters, and switching the motion mode between wheeled fast driving, low-speed wheeled driving, wheel-leg coordinated obstacle crossing and safe stopping according to the terrain score, the problem of insufficient obstacle crossing ability of wheeled robots in grass, mud and ditch sections and the problem of limited speed and endurance of legged robots on long-distance flat roads can be improved, which is conducive to maintaining continuous long-distance boundary patrol.
[0045] In one embodiment, the airport perimeter segment map includes multiple fence segments; each fence segment has a unique number, and each fence segment is associated with fence information and area information, namely the fence segment number, route coordinates, location coordinates, and unified timestamp, which are stored together.
[0046] In the above implementation process, by configuring a unique number for each fence segment and associating it with geometric, benchmark, calibration and communication attributes, positioning, touch, vision, vibration, alarm and communication data are bound to the same spatiotemporal index, which can improve the problem of data from different sources being difficult to correlate with each other, and is conducive to the traceability and verification of inspection records.
[0047] In one embodiment, the method further includes: in the task initialization state, loading the normal appearance reference, mesh geometric reference, connection point reference, sign position and attitude reference, perimeter box external closure reference, pole and box external geometric reference, basic neighborhood reference, normal fence response reference and applicable parameter version according to the fence segment number; projecting the current positioning coordinates to the target fence segment or facility route coordinates, and forming an observation pose consistent with the fence normal, inspection offset line and safety distance; maintaining a non-contact detection mode of vision, thermal imaging, lidar and passive vibration when the position confidence, lateral distance uncertainty, fuselage attitude, lateral space, support margin or target safety distance do not meet the requirements; for the same fence segment The system can collect visible light images, infrared or thermal imaging images, laser point clouds, inertial measurement attitude, fence vibration signals, and operating status from multiple observation poses, and register the laser point clouds with the corresponding geometric references, outputting the registration results and registration quality marks. If candidate anomalies are detected in visual, thermal imaging, laser point cloud, or passive vibration, and the safety conditions are met and there is no risk of personnel, animals, or unknown targets near the fence, the side touch detection module is invoked to verify the safe-to-access detection points. The inspection results of the fence and its ancillary facilities are bound to the task number, event category, fence segment number, route coordinates, geographical coordinates, unified timestamp, sensor source, input quality mark, reference version, parameter package version, handling status, and recording status.
[0048] In the above implementation process, by binding the inspection results of the fence and its ancillary facilities to the fence section number, benchmark version, parameter package version and input quality mark, and limiting active touch control to a verification method only when safety conditions are met and there are no people, animals or unknown targets near the fence, structural condition evidence can be obtained without increasing the risk of contact. This is conducive to making the external condition conclusions of the fence body, poles, facilities and foundation traceable and verifiable.
[0049] In one embodiment, the fence and its ancillary facilities include a fence body. The inspection items for the fence body include at least one of the following: fence damage inspection, fence corrosion, fallen fence screws, foreign objects hanging from the fence, sign stability, other appearance abnormalities, and fence perimeter enclosure. The fence damage inspection is triggered when the visible light image shows abnormal grid continuity, linear gaps, or disconnection at joints; the laser point cloud shows fence outline shift, partial missing parts, or abnormal openings; or passive vibration causes abnormal impacts or continuous vibrations. Image features are correlated with point cloud features, and output for verification when only a single modality exists and occlusion, shadows, vegetation, reflections, or missing point clouds cannot be ruled out. The fence corrosion inspection is triggered when the current visible light texture, color, surface mottled features, or changes in the appearance of joints are abnormal compared to the normal appearance benchmark of the same fence type and environmental range. The touch result is only used as a structural status indicator. Supplementary evidence is used, and touch results are not used to replace the judgment of rust appearance. The inspection of fallen fence screws is triggered when the visible appearance of the connection point is inconsistent with the reference of the connector, the relative position of adjacent components changes, or there is a local geometric anomaly in the point cloud. The inspection of fallen connectors, loose connectors, and general obstruction are distinguished by equivalent stiffness, normalized hysteresis, residual displacement, and vibration characteristics. The inspection of foreign object suspension in the fence is triggered when the visible object still maintains a spatial relationship with the fence mesh or signage neighborhood under multiple observation poses. When the target is in a continuous moving state and the target trajectory crosses the center line judgment surface of the fence from the outside to the inside, the inspection switches to the identification of airborne object intrusion. The inspection of the fence perimeter box only identifies the outer contour of the box, the door cover plane, the closed boundary, the tilt state, the connection relationship of the surrounding fence, the basic neighborhood, and visible external damage. The box is not opened, the internal electrical components are not touched, and the internal electrical status is not read.
[0050] In one embodiment, the fence and ancillary facilities include monitoring poles, light poles, intermediate distribution boxes, and distribution boxes within the monitoring station. The complete inspection of the monitoring poles is used to identify any missing, tilted, sagging, or obviously detached components of the pole body, equipment suspension positions, speakers, supplementary lighting, audible and visual alarms, lightning rods, and their visible external installation status. The complete inspection of the light poles is used to identify abnormalities in the pole's centerline, light position, light posture, light sagging status, pole tilt, and relative foundation position, without interpreting the absence of thermal response solely as a lighting failure or missing component. The inspection of the intermediate distribution boxes and the distribution boxes within the monitoring station is only used to identify the external box outline, door closure status, box tilt, water accumulation near the foundation, and relative foundation position. The boxes are not opened, internal wiring is not touched, or the internal electrical status is not read. External temperature maps, door appearance, or foundation status are not used to presume internal power distribution faults. In cases where equipment is obstructed, point cloud registration is insufficient, nighttime image quality is substandard, or positioning uncertainty affects facility attribution, the status of the corresponding facility is output for verification.
[0051] In one embodiment, the method further includes: identifying visible water accumulation candidates around low-lying areas, box foundations, and pole foundations; recording quality defects where water surface reflection, absorption, occlusion, or sparse point clouds lead to unreliable laser results, without interpreting missing point clouds as water accumulation; forming water accumulation candidates when multi-view visual or thermal imaging evidence is consistent with the low-lying area baseline and the point cloud or ground geometry supports the corresponding positional relationship; otherwise, outputting the low-lying area status for verification; identifying continuous geometric offsets of the edges, adjacent columns, mesh baselines, and ground elevation relationships of the fence foundations, pole foundations, and box foundations relative to historical baselines; forming foundation settlement candidates when continuous and unidirectional geometric offsets are present and the registration quality, positioning quality, and baseline matching conditions meet the requirements; wherein, the effective displacement, equivalent stiffness, normalized hysteresis, vibration energy, vibration attenuation coefficient, residual displacement, and spatial consistency obtained by active touch control serve as auxiliary structural evidence; and collecting long-distance evidence while maintaining a safe distance when local ground accessibility does not meet the requirements.
[0052] In one embodiment, the method further includes: loading the route coordinate range, permissible boundaries, test direction, road segment baseline, drivability score parameter package, and safe stopping strategy of the test segment in the task initialization state; and writing the actual road segment width, slope range, effective length, surface condition, load-bearing diameter, total robot test mass, environmental conditions, test direction, parameter version, and expected passage method into the test record; using LiDAR to establish a local elevation map, and combining data from visible light, infrared or thermal imaging, inertial measurement unit, wheel speed sensor, and foot contact force sensor to obtain input evidence of slope, roughness, ditch width and depth, wheel-ground adhesion status, fuselage stability margin, footing feasibility, and slippage rate; reducing the credibility of the drivability score for ground attributes that cannot be reliably estimated by existing sensor quality; and when the drivability score meets the conditions for high-speed wheeled driving and the slippage rate, The vehicle maintains wheeled driving when attitude and boundary constraints meet requirements; switches to low-speed wheeled driving when the slope, roughness, or slippage rate increases but continuous rolling conditions still hold; switches to wheel-foot coordination mode when ditches, steps, insufficient continuous rolling, or more reliable landing plans exist; switches to safe stop when wheel-ground adhesion, vehicle stability margin, landing feasibility, positioning quality, lateral boundary, or sensor quality do not meet safety conditions; the switching adopts a hysteresis rule; if the vehicle completes the prescribed actions of entering, continuous passage, controlled deceleration or braking, and leaving the road section within the test direction and the permitted passage boundary, and does not trigger stability margin, lateral boundary, continuous slippage, critical sensor failure, or safety control overstepping conditions, it is recorded as passing; if the road conditions exceed the calibration range or the sensor quality is insufficient, a safe stop or no valid conclusion is selected.
[0053] In the above implementation process, by writing the road width, slope, load-bearing diameter and total test mass into the test record and clearly stating that "failure to pass must not be interpreted as safe passage", it is possible to avoid misinterpreting safe stopping or detour as passing the test, which helps to make the passing test conclusion correspond to the actual road conditions.
[0054] In one embodiment, the method further includes: establishing scenario-based functional verification records according to the scenario, time pattern, detection object, expected output, and handling status; and outputting no valid conclusion or pending verification in cases where there is a lack of effective normal baseline, the current environment is an out-of-sample environment, the sensor quality does not meet the conditions, the parameter package is expired, the model version is mismatched, or the event location cannot be reliably bound to the fence segment, while saving the available original evidence.
[0055] In the above implementation process, by distinguishing the conclusion status into confirmed, candidate pending verification, no valid conclusion and safety suspension, and stipulating that non-confirmed status cannot be described as completed confirmation, it is possible to avoid describing the result of insufficient evidence as the fence being normal, which is conducive to the consistency between the inspection report and the actual evidence status on site. Secondly, embodiments of this application also provide a perimeter inspection robot, comprising: a robot body, a robot controller, a positioning module, a touch detection module, an environmental perception module, an intrusion recognition module, a fusion decision module, and a communication module. The robot controller is connected to the positioning module, the touch detection module, the environmental perception module, the intrusion recognition module, the fusion decision module, and the communication module. The positioning module is used to determine the fence segment number where the perimeter inspection robot is located and the positioning coordinates of the perimeter inspection robot based on a pre-established airport perimeter segment map and a unified time reference. The robot controller is used to control the perimeter inspection robot to enter a safe approach state and adjust its body posture when safety conditions are met, and to synchronously collect the current action parameters of the perimeter inspection robot. The touch detection module is used to determine the effective displacement of the fence along the loading direction based on the current action parameters. The environmental perception module is used to collect image information and determine the fence intrusion target. The target detection results; the environmental perception module is also used to collect image information of the area inside and outside the fence, perform target recognition on the image information, determine the target category of the fence intrusion target, and establish the target trajectory of the fence intrusion target; the intrusion recognition module is used to establish a fence centerline determination surface based on the fence centerline and the inner and outer directions of the fence segment associated with the fence intrusion target, and determine whether the fence intrusion target is located outside the fence, in the fence neighborhood, inside the fence, or has made a valid crossing from the outside to the inside based on the spatial relationship between the target trajectory and the fence centerline determination surface, and output personnel and biological intrusion recognition results; the fusion decision module is used to associate the physical inspection results corresponding to the fence defect category with the personnel and biological intrusion recognition results according to the fence segment number and time window, and determine the alarm level based on the confirmed intrusion, physical state, or security risk event; the communication module is used to upload the alarm evidence package and inspection record to the airport inspection management platform according to the alarm level.
[0056] In the above implementation process, by making the outputs of each module, such as positioning, touch control, perception, fusion decision and communication, become the trigger conditions or constraints of the next module in sequence, the functional units of the robot form a collaborative closed loop rather than simply running in parallel. This can improve the problem that each module can operate independently and cannot verify each other, and is conducive to improving the consistency and traceability of the perimeter inspection results.
[0057] In one embodiment, the robot body is a multi-legged wheeled body; the multi-legged wheeled body includes a body, multiple mechanical legs, and a drive wheel disposed at the end of each mechanical leg.
[0058] In the above implementation process, by adopting a wheel-foot composite body consisting of a body, multiple mechanical legs and an end-drive wheel, the robot can not only travel continuously on flat inspection roads in a wheeled manner, but also pass through ditches, slopes and uneven areas in a wheel-foot coordinated manner, which is conducive to balancing long-distance inspection efficiency and the ability to pass through complex terrain.
[0059] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, specific embodiments are described below in conjunction with the accompanying drawings. Attached Figure Description
[0060] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0061] Figure 1 This is a schematic diagram of the perimeter inspection robot provided in an embodiment of this application; Figure 2 This is a schematic diagram illustrating the overall architecture and communication relationships of the perimeter inspection robot provided in an embodiment of this application. Figure 3 A flowchart of an airport perimeter inspection method provided in an embodiment of this application; Figure 4 A schematic diagram illustrating autonomous airport perimeter patrol, target identification, and alarm verification provided in this application embodiment; Figure 5 This is a schematic diagram of the wheel-foot composite motion and terrain adaptation process provided in the embodiments of this application; Figure 6 This is a schematic diagram of the retracted body of the perimeter inspection robot provided in an embodiment of this application; Figure 7 This is a schematic diagram of the unfolded body of the perimeter inspection robot provided in an embodiment of this application.
[0062] Figure descriptions: ①- Positioning module, ②- Robot controller, ③- Touch detection module, ④- Environmental perception module, ⑤- Communication module. Detailed Implementation
[0063] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.
[0064] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.
[0065] With the expansion of civil airport operations, airport perimeter fencing, as a physical barrier to prevent unauthorized personnel, animals, and other targets from entering the airport control area, has received increasing attention regarding its inspection coverage and quality. Airport perimeter fencing is continuously deployed along the airport's perimeter, covering a long distance and a wide area, and traversing patrol roads, grasslands, muddy ground, drainage ditches, slopes, and uneven terrain. The condition of the perimeter fencing, surrounding targets, and inspection road conditions are constantly changing, and inspection work is also affected by environmental factors such as nighttime, rain, snow, fog, haze, strong winds, and low light levels.
[0066] Currently, airport perimeter inspections mainly rely on manual observation of the mesh panels, posts, and connectors, with personnel manually shaking the mesh when they suspect it is loose or unstable. This method is labor-intensive, has long inspection cycles, and the judgment results depend on personnel experience. It is difficult to generate structured records that are associated with fence segment numbers, accurate coordinates, detection waveforms, and algorithm versions, and it is also difficult to maintain continuous inspections in inclement weather or remote perimeter sections.
[0067] Fixed cameras, perimeter vibration sensors, and ordinary mobile inspection equipment can provide some images or alarm information, but problems still exist, such as coverage blind spots, false alarms caused by wind blowing vegetation and changes in shadow, difficulty in quickly verifying alarm points, and inconsistencies in the spatiotemporal reference of data from different sensors. Visual results alone are insufficient to directly confirm whether the mesh is loose, wires are broken, welds are detached, permanently deformed, or the column is unstable; contact force thresholds alone are also insufficient to distinguish between localized mesh loosening and low-frequency swaying of the supporting structure.
[0068] In terms of mobile platforms, airport patrol roads include smooth surfaces suitable for rapid rolling, as well as grass, mud, ditches, and uneven areas. Traditional wheeled robots lack obstacle-crossing capabilities, and ordinary legged robots are limited in speed and range on long distances of smooth surfaces; satellite positioning may be degraded in areas with buildings, trees, or complex electromagnetic environments, and single positioning methods are insufficient to stably control the distance and attitude between the robot and the fence.
[0069] Therefore, how to conduct continuous, quantifiable, and traceable inspections of airport perimeters under complex terrain and day-night conditions, and how to distinguish between different perimeter events such as crowds gathering on the perimeter, people climbing, people crossing, continuous stay on the inner perimeter, airborne objects crossing the boundary, foreign objects hanging, and authorized operations, are urgent problems to be solved.
[0070] In view of this, this application proposes an airport perimeter inspection method. By organizing the positioning, approach, detection, diagnosis, alarm and recording stages using a segmented map of the airport perimeter and a unified time reference, the output of the previous stage becomes the trigger condition or constraint condition of the next stage. The spatial relationship between the target and the perimeter is transformed into a determinable motion state by using the perimeter centerline determination surface. This can improve the problem that multiple detection modules operate independently and are difficult to verify each other, and is conducive to forming a traceable perimeter inspection closed loop.
[0071] To facilitate understanding of this embodiment, a detailed description of a perimeter inspection robot that performs a perimeter inspection method disclosed in this application embodiment will be provided first.
[0072] like Figure 1 The diagram shown is a structural schematic of the perimeter inspection robot provided in this application embodiment, including: robot body, robot controller ②, positioning module ①, touch detection module ③, environmental perception module ④, fusion decision module and communication module ⑤. The robot controller ② is connected to the positioning module ①, touch detection module ③, environmental perception module ④, intrusion recognition module, recording module and communication module ⑤ respectively.
[0073] The system comprises the following modules: Positioning module ① determines the fence segment number and coordinates of the perimeter inspection robot based on a pre-established airport perimeter segment map and a unified time reference; Robot controller ② controls the perimeter inspection robot to enter a safe approach state and adjust its posture when safety conditions are met, and simultaneously collects the robot's current motion parameters; Touch detection module ③ determines the effective displacement of the fence along the loading direction based on the current motion parameters; Environmental perception module ④ collects image information and determines the detection result of fence intrusion targets; Fusion decision module determines the fence state feature quantity based on the effective displacement, outputs the fence defect category based on the deviation between the fence state feature quantity and the normal fence response benchmark corresponding to the fence segment number, and, according to the fence segment number and time window, compares the diagnostic results corresponding to the fence defect category with the fence intrusion criteria. The detection results of the target are correlated and fused to obtain the alarm confidence level, and the alarm level is determined based on the alarm confidence level; the intrusion identification module is used to establish the fence centerline judgment surface based on the fence centerline and the inner and outer directions of the fence segment associated with the fence intrusion target, and to determine whether the fence intrusion target is outside the fence, in the fence neighborhood, inside the fence, or has made a valid crossing from the outside to the inside based on the spatial relationship between the target trajectory and the fence centerline judgment surface, and outputs the personnel and biological intrusion identification results; the fusion decision module is used to correlate the physical inspection results corresponding to the fence defect category with the personnel and biological intrusion identification results according to the fence segment number and time window, and to determine the alarm level based on the confirmed intrusion, physical state or security risk event; the recording module is used to record the alarm evidence package and inspection record according to the alarm level; the communication module is used to upload the alarm evidence package and inspection record to the airport inspection management platform.
[0074] The output of the preceding module here serves as the triggering condition or constraint condition for the subsequent module.
[0075] Understandably, by having each module sequentially receive the output of the previous module and using that output as the trigger or constraint condition for this module, the robot's internal positioning, approach, detection, diagnosis, fusion, and uploading functions can form a collaborative relationship rather than a parallel relationship, thus improving the problem that it is difficult for each functional unit to verify each other when operating independently.
[0076] In one embodiment, such as Figure 2 As shown, the robot controller ② is connected to a visible light camera, infrared or thermal imaging equipment, lidar, IMU, wheel speed sensor, foot contact force sensor, force and displacement sensors of the touch actuator, fence vibration sensor and communication module ⑤ via Ethernet or internal data bus.
[0077] By connecting various sensors and communication terminals through the robot controller ② and using the internal bus to carry synchronous acquisition commands, multi-channel data can be aligned under a unified time reference, providing the conditions for effective displacement compensation and multimodal correlation.
[0078] In the above implementation process, by making the outputs of each module, such as positioning, touch control, perception, fusion decision and communication, become the trigger conditions or constraints of the next module in sequence, the functional units of the robot form a collaborative closed loop rather than simply running in parallel. This can improve the problem that each module can operate independently and cannot verify each other, and is conducive to improving the consistency and traceability of the perimeter inspection results.
[0079] In one embodiment, the robot body is a multi-legged wheeled body.
[0080] The multi-legged wheeled body includes a body, multiple mechanical legs, and a drive wheel located at the end of each mechanical leg.
[0081] By equipping each mechanical leg with a drive wheel, the robot can continuously roll on flat surfaces while retaining the lifting and landing capabilities of the mechanical legs for obstacle crossing. This improves the adaptability of single-wheeled or single-legged vehicles to complex terrain within airport perimeters.
[0082] Optionally, the multi-legged wheeled body has four mechanical legs, each with a drive wheel at its end. Sensing, positioning, touch control, and communication devices are arranged on the top or side of the body. The number of mechanical legs mentioned above is merely exemplary and can be selected according to actual needs.
[0083] Please see Figure 1 , Figure 1 This is an isometric schematic diagram of a multi-legged wheeled robot provided in an embodiment of this application. The robot uses a multi-legged wheeled body as its moving carrier, with drive wheels at the ends of its four mechanical legs. Sensing, positioning, touch control, and communication devices are arranged on the top or sides of the robot body and coordinated by a unified controller. The structure shown is used to illustrate the relative configuration of functional modules; the specific shape, mounting base, and housing can be adjusted according to the protection level and load.
[0084] By centralizing the deployment of sensing, positioning, touch, and communication devices and coordinating them with a unified controller, the timing differences caused by the separate wiring of multiple devices are reduced, which is conducive to multi-channel synchronous acquisition under a unified clock.
[0085] In the above implementation process, by adopting a wheel-foot composite body consisting of a body, multiple mechanical legs and an end-drive wheel, the robot can not only travel continuously on flat inspection roads in a wheeled manner, but also pass through ditches, slopes and uneven areas in a wheel-foot coordinated manner, which is conducive to balancing long-distance inspection efficiency and the ability to pass through complex terrain.
[0086] The perimeter inspection robot in this embodiment can be used to execute various steps in the methods provided in the embodiments of this application. The implementation process of the airport perimeter inspection method is described in detail below through several embodiments.
[0087] Please see Figure 3 This is a flowchart of the airport perimeter inspection method provided in the embodiments of this application. The following will describe... Figure 3 The specific process shown will be explained in detail.
[0088] It should be noted that the threshold, weight, sampling frequency, filtering parameters, frequency band, inspection distance, contact force, and associated time window in the embodiments of this application are all configured according to the airport fence structure, road conditions, calibration data, and safety management requirements.
[0089] Step S201: Based on the pre-established airport perimeter segment map and unified time reference, determine the fence segment number where the perimeter inspection robot is located and the positioning coordinates of the perimeter inspection robot.
[0090] In one embodiment, the wheel and foot status, positioning results, raw touch data and diagnostic results, and visible light or infrared sensing results of the perimeter inspection robot are respectively connected to the robot controller ②. The airport inspection management platform issues inspection tasks, fence segment maps, parameter versions, and handling instructions, and the robot uploads its status, defect results, intrusion alarms, images and videos, waveform summaries, and audit logs.
[0091] By unifying the data from each functional module of the perimeter inspection robot into the robot controller ② and the communication terminal, and by having the platform uniformly distribute the perimeter segment map and parameter version, a two-way controllable data channel is formed between the perimeter inspection robot and the platform, which is conducive to keeping the perimeter segment number, parameter version and uploaded evidence consistent.
[0092] Optionally, such as Figure 4 As shown, the platform divides the fence into fence segments with unique numbers and sets closed-loop or round-trip routes, permitted access areas, prohibited areas, charging points, and key detection points. The perimeter inspection robot patrols along the inner side of the fence; when it identifies a target, receives a fixed perimeter alarm, or detects an anomaly in touch diagnostics, it projects the event coordinates onto the fence segment and route coordinates, plans a safety verification path, and updates the event evidence.
[0093] By projecting event coordinates onto fence segments and route coordinates and planning verification paths accordingly, autonomous navigation, target recognition, and alarm verification can share the same spatial index, which can improve the problem of alarm points being difficult to verify quickly.
[0094] Step S202: Under the condition that the safety conditions are met, control the perimeter inspection robot to enter the safe approach state and adjust the body posture; and simultaneously collect the current motion parameters of the perimeter inspection robot.
[0095] The current action parameters include: contact force, displacement of the side touch detection module ③, and base displacement.
[0096] In one embodiment, after acquiring the current action parameters, the raw signals such as the current action parameters can be preprocessed first.
[0097] Optionally, the preprocessing may include methods such as anti-aliasing filtering, outlier removal, time alignment, and uniform sampling rate resampling. The preprocessing method can be selected according to the actual situation.
[0098] In another embodiment, zero bias can be estimated using the uncontacted area, with the contact interval starting when the contact force exceeds the contact establishment threshold and the duration meets the condition, and ending when the force falls back after unloading; the contact establishment threshold and duration are generated by parameter samples such as the current action parameters of the unloaded and normal fence calibration.
[0099] Understandably, by resampling at a uniform sampling rate and determining the contact interval using the calibrated contact establishment threshold and duration, the effective displacement calculation can be avoided by mixing the contact establishment segment and the unloading segment, which is beneficial to improving the robustness of subsequent fitting.
[0100] Step S203: Determine the effective displacement of the fence along the loading direction based on the current action parameters.
[0101] Step S204: Determine the fence state characteristic quantity based on the effective displacement, and output the fence defect category based on the deviation between the fence state characteristic quantity and the normal fence response benchmark corresponding to the fence segment number.
[0102] Step S205: Collect image information of the area inside and outside the fence, perform target recognition on the image information, determine the target category of the fence intrusion target, and establish the target trajectory of the fence intrusion target; establish the fence centerline determination surface based on the fence centerline and the inner and outer directions of the fence segment associated with the fence intrusion target; determine whether the fence intrusion target is located outside the fence, in the fence neighborhood, inside the fence, or has made a valid crossing from the outside to the inside based on the spatial relationship between the target trajectory and the fence centerline determination surface, and output the personnel and biological intrusion identification results.
[0103] Step S206: According to the fence segment number and time window, associate the physical inspection results corresponding to the fence defect category with the personnel and biological intrusion identification results, and determine the alarm level based on the confirmed intrusion, physical status or security risk event.
[0104] Step S207: According to the alarm level, upload the alarm evidence package and inspection record to the airport inspection management platform.
[0105] The robot controller ② operates a closed-loop task state machine, dividing the task process into M0 task initialization, M1 boundary localization, M2 safe approach, M3 touch or passive detection, M4 state diagnosis, M5 personnel and biological intrusion identification or physical state result processing, M6 event classification alarm, M7 data recording, and M8 inspection recovery or safety handling status. The states are not sequentially linked in a fixed order, but are driven by location confidence, target events, touch safety conditions, diagnostic quality, and task strategy.
[0106] In this system, the fence segment number, route coordinates, and confidence interval output by M1 serve as inputs for M2 to plan the touch pose. If the confidence level decreases, M2 will not allow further approach and will return to M1 for repositioning. The fence normal, lateral distance, and stable attitude output by M2 are used for the actuator direction and load limit of M3. If safety conditions are not met, M3 will switch to collecting only visual, thermal imaging, and passive vibration data. The multi-channel synchronous data and quality markers of M3 are input to M4. M4 outputs the physical status diagnosis results of the fence or facility, the comprehensive anomaly score, and the validity marker. The environmental perception module synchronously outputs the target trajectory, inner and outer markings, climbing or crossing features, and recognition quality of personnel, animals, or airborne objects. M5 organizes the target recognition results, physical status results, and items to be reviewed according to the fence segment number, route coordinates, and unified timestamp, and can trigger M2 to replan the review perspective or detection point. The event level, evidence summary, and handling status generated by M6 are input to M7. M7 saves event numbers, locations, key images, video indexes, waveform summaries, input quality, and original evidence indexes according to the task strategy; after recording is completed or the local persistence conditions are met, the system enters M8 and continues inspection; when positioning, attitude, power, key sensors are continuously abnormal or safety constraints are continuously exceeded, M8 executes stop, return to home, or remote takeover.
[0107] By making each state driven by confidence, events, security conditions, and task policies rather than in a fixed unidirectional chain, the output of the previous state becomes the triggering condition, constraint condition, or evidence of the next state, which can improve the problem of multiple independent modules being simply parallel and difficult to verify each other.
[0108] In one embodiment of this application, the airport perimeter inspection method is executed according to the following complete process: The airport inspection management platform generates a task and issues a set of perimeter segments, inspection sequence, permitted access areas, prohibited areas, key detection points, parameters and map versions, business strategies, and completion deadlines. The airport perimeter inspection robot completes identity authentication, sensor self-test, unified clock calibration, positioning initialization, touch-screen idle check, and task execution condition check; if the conditions are not met, the reason is recorded and related functions are disabled. The airport perimeter inspection robot selects a wheeled or wheel-footed cooperative motion mode based on the terrain score, continuously integrates positioning information, and cruises along the inner side of the perimeter, projecting its position onto the perimeter segment and route coordinates. Before reaching the detection point, it assesses the positioning reliability, perimeter distance, fuselage attitude, support margin, and lateral space; if the conditions are met, it enters a safe approach; if not, it repositions, adjusts its posture, or skips active touch. It executes preset touch actions, simultaneously collecting force, displacement, base movement, vibration, IMU, and drive status; if the safety constraints are exceeded, it immediately releases force and retracts. The system performs attitude and mechanism flexibility compensation, calculates equivalent stiffness, hysteresis loop area, wavelet packet energy, attenuation coefficient, residual displacement, repeatability, and baseline deviation, and outputs diagnostic probability and data quality. It establishes continuous target trajectories for personnel, animals, and airborne objects, and outputs results for crowd attention, personnel climbing, personnel or biological intrusion, airborne object boundary crossing, authorized personnel verification, or pending verification based on the fence centerline, inner and outer markings, fence distance, climbing characteristics, crossing characteristics, and recognition quality. Based on confirmed personnel or biological intrusion, personnel climbing, high-intensity impact, abnormal fence physical state, passage risk, and task safety rules, it generates observation, attention, warning, or emergency alarms, forming an evidence package containing fence segment number, location, time, model version, and input quality; new perspectives or detection points are planned when verification is needed. The robot records event summaries, location status, key images, video indexes, waveform summaries, and quality markers according to the task strategy, forming a traceable inspection record, which is uploaded to the airport inspection management platform by the communication module. After completing the target fence segment, the inspection record is updated. If the battery is low, the positioning is continuously abnormal, the critical sensors fail, or the safety constraints are continuously exceeded, the system will enter a safe stop, return to base, or remote takeover phase. After the task is completed, the platform generates a status report for each fence segment, including the defect location, category, severity, evidence quality, intrusion event, review conclusion, robot trajectory, model and map version, and a list of robots to be repaired.
[0109] Understandably, this involves task assignment, self-check, cruise control, proximity detection, touch control, diagnostics, fusion, and alarms. Understandably, by linking task issuance, self-inspection, patrol, approach, touch control, diagnosis, identification, alarm, recording and task completion into a single process, and setting back branches at each stage when conditions are not met, the entire perimeter inspection process becomes traceable, interruptible and recoverable, which is beneficial for maintaining continuous inspections on perimeters over long distances and in various environmental conditions.
[0110] In the above implementation process, by organizing the positioning, approach, detection, diagnosis, identification, alarm and recording stages using the segmented map of the airport perimeter and a unified time reference, the output of the previous stage becomes the trigger condition or constraint condition of the next stage. The spatial relationship between the target and the perimeter is transformed into a determinable motion state by using the perimeter centerline determination surface. This can improve the problem that multiple detection modules operate independently and are difficult to verify with each other, and is conducive to forming a traceable perimeter inspection closed loop.
[0111] In one possible implementation, a fence centerline determination surface is established based on the fence centerline and inner / outer directions of the fence segment associated with the fence intrusion target. This includes: obtaining the fence centerline, inner / outer directions, and allowed association range of the fence segment associated with the fence intrusion target stored in the airport boundary segment map, and establishing the fence centerline determination surface based on the fence centerline and inner / outer directions.
[0112] In one embodiment, when the fence intrusion target is located outside the fence and moves parallel to it, an outer trajectory record is formed; when the fence intrusion target continues to approach the fence, an approach trend record is formed.
[0113] The target trajectory here records the continuous changes in the position, speed, direction of movement, and closest distance to the center line of the fence intrusion target, which is used for subsequent climbing, crossing, crowd gathering, and intrusion identification.
[0114] The aforementioned fence centerline determination surface, inner and outer markings, and continuous trajectory are only used for identifying spatial relationships within the fence and do not directly lead to intrusion conclusions based on single-frame target categories, single thermal image contours, single vibrations, or single point cloud results.
[0115] Among them, the fence centerline determination surface is used to determine whether the fence intrusion target is outside the fence, in the fence neighborhood, inside the fence, or has made a valid crossing from the outside to the inside.
[0116] In the above implementation process, the fence centerline determination surface is established by storing the fence centerline, inner and outer directions and allowed association range in the airport perimeter segment map. The spatial state of the target is quantified into outer trajectory records, approach trend records and trajectory basis for climbing, crossing, gathering and intrusion identification. This avoids making intrusion conclusions directly based on single frame images, single thermal image contours, single vibrations or single point clouds, which helps to improve the verifiability of intrusion identification and reduce false alarms.
[0117] In one possible implementation, the method further includes identifying airborne object intrusions: establishing continuous trajectories for airborne targets in the areas inside and outside the fence, and projecting these trajectories onto the fence centerline determination surface, fence segments, and the controlled area inside the fence boundary; generating an airborne object intrusion identification result when an airborne target enters the fence neighborhood from the outside and continues to move inward, or enters the controlled area and maintains a traceable trajectory; and outputting an unidentified airborne target if the category confidence of the airborne target is insufficient. If an airborne target is only suspended on fence mesh, signs, or connecting components and does not form a continuous boundary-crossing trajectory from the outside to the inside, the airborne target is classified as a foreign object suspension inspection and is not treated as an airborne object intrusion.
[0118] Among them, airborne targets include at least one of clothing, electrical wires, garbage bags, and other airborne objects.
[0119] In one embodiment, for an airborne target, visible light is used to identify the target category and motion profile, lidar is used to assist in estimating the spatial relationship between the airborne target and the fence, and infrared or thermal imaging is used to assist in distinguishing the airborne target from the background thermal profile.
[0120] In the above implementation process, by establishing a continuous trajectory for the airborne target and projecting it onto the judgment surface, fence section, and controlled area, the airborne object intrusion identification result is output only when a traceable trajectory is formed from the outside to the inside or when it enters the controlled area. Airborne targets that are only suspended and have not crossed the boundary are classified as foreign object suspension inspections, avoiding direct false alarms of fence attachments as intrusions. At the same time, the multimodal combination of visible light, lidar, and infrared or thermal imaging improves the quality of airborne target category and spatial relationship identification.
[0121] In one possible implementation, the method further includes the identification of nearby crowds: the trajectories of multiple individuals located in the vicinity outside the fence are spatiotemporally aggregated, and the number of individuals, spatial density, dwell time, consistency of movement direction, changes in distance from the center line of the fence, and approach trends of the multiple individuals are analyzed; when multiple individuals do not approach the fence, do not exhibit climbing characteristics, do not exhibit crossing characteristics, and are not effectively associated with abnormal vibrations in the same fence segment, a crowd attention event is output, and no illegal intrusion alarm is directly triggered; when any one of the multiple individuals continuously approaches the fence, exhibits climbing characteristics, exhibits crossing characteristics, or is associated with abnormal vibrations, an intrusion identification event is re-established centered on the trajectory of that individual.
[0122] Among them, the crowd discrimination parameters are generated from airport scenario calibration samples, mission safety rules or parameter packages.
[0123] In the above implementation process, by spatiotemporally aggregating the trajectories of multiple people in the outer neighborhood and distinguishing between crowd attention events and intrusion identification events, the security alarms for group gatherings are retained, while avoiding directly treating the gathering of people who have not approached, climbed, or crossed as illegal intrusions.
[0124] In one possible implementation, the method further includes personnel climbing identification: when the outline of the fence intrusion target continuously overlaps with the fence area, the target trajectory continuously changes along the fence height direction and conforms to climbing characteristics, and the fence vibration sensor of the same fence segment or adjacent fence segments allowed by calibration outputs abnormalities within the associated time window, a climbing probability and identification quality mark of the fence intrusion target are formed, and the trajectory identifier, fence segment number, route coordinates, shortest distance, fence overlap continuity, height direction change characteristics, movement direction characteristics, vibration correlation and input quality of the fence intrusion target are recorded.
[0125] If the fence vibration quality is substandard, the fence intrusion target is severely obstructed, or the inner and outer markings are unstable or the trajectory is not continuous, the personnel climbing will be sent for verification. If the fence intrusion target is a person and is located near the fence, the fence inspection robot will be controlled to maintain a safe distance for observation, without approaching, loading, or contacting the fence; after the personnel leave and the safety conditions are restored, the physical condition of the fence will be checked.
[0126] In the above implementation process, the overlapping of contours, continuous changes in height direction and abnormal vibration of vibration sensors in the same or adjacent fence segments are used as climbing judgment conditions. When the vibration quality, obstruction or trajectory continuity is insufficient, the system outputs that the person is climbing and needs to be verified, thus avoiding the identification of climbing based on a single image clue. At the same time, the robot is prohibited from approaching, loading or touching the fence when the person is near it, and physical verification is arranged only after the person leaves and the safety conditions are restored, which helps to ensure personal safety.
[0127] In one possible implementation, the method further includes personnel and biological intrusion identification: The target trajectory is continuously checked before, during, and after crossing to exclude apparent crossings caused by temporary obstruction, viewpoint switching, location jumps, or unstable fence segment affiliation; when the target trajectory continuously enters the fence from the outside to the inside, or the fence intrusion target remains within the fence boundary and effectively crosses the fence centerline determination plane, the personnel intrusion identification result or biological intrusion identification result is output, and the relevant trajectory, inside / outside markers, crossing time, fence segment number, route coordinates, geographic coordinates, and quality markers are recorded; when the fence intrusion target only briefly enters the fence neighborhood, does not effectively cross the fence centerline determination plane, cannot be continuously tracked, or the evidence quality is insufficient, a pending verification result is output, and the fence intrusion target is not identified as a definitive intrusion.
[0128] Among them, dogs, birds and other animals form biological invasion identification results based on the relationship between the inside and outside of the trajectory and the effective crossing relationship.
[0129] In the above implementation process, the continuity of the trajectory before, during and after crossing is checked to eliminate the apparent crossing caused by temporary occlusion, viewpoint switching, and positioning jump. The continuous entry into the inner side, the continuous location in the inner side and the effective crossing are used as combination conditions to improve the credibility of personnel and biological intrusion identification results and avoid identifying targets that enter the neighborhood for a short time or do not cross effectively as deterministic intrusions.
[0130] In one possible implementation, the method further includes authorized operator verification: based on the visible appearance information, trajectory, location, and current time of the fence intrusion target, consistency verification is performed with the valid work permit information issued for the current task to generate an authorized operator verification result; if the fence intrusion target is located within the permitted work area and valid time period, the visible appearance information is consistent with the parameter package, and the valid work permit information matches the current task, the authorized operator verification is output as passed; if the valid work permit information is valid but the appearance features, trajectory continuity, visibility ratio, observation distance, or nighttime imaging quality are insufficient, the authorized operator is output as pending review; if the fence intrusion target is located in a non-permitted area, the work time period is inconsistent, the permit status is invalid, or the visible evidence is insufficient, the personnel identity is output as pending verification, and the personnel intrusion identification and event classification rules continue to apply.
[0131] The visible appearance information includes at least one of the following: workwear appearance, reflective vest, and safety helmet.
[0132] The valid work permit information here includes at least the task association, permitted work area, valid time period, work type, and permit status.
[0133] The above consistency verification does not read personal identity information or perform facial recognition, and does not use thermal image contours as the sole basis for identity verification.
[0134] In the above implementation process, by verifying the consistency between the appearance of the work clothes, reflective vests, safety helmets, personnel trajectory, location, current time and valid work permit information, the system outputs "Authorized operator verified," "Authorized operator pending review," or "Personnel identity pending verification." It does not read personal identity information or perform facial recognition, which can meet the needs of operator management while avoiding personal identity verification.
[0135] In one possible implementation, the method also includes day / night mode and recognition quality: the current day / night mode is determined based on environmental perception quality, a unified time reference, and the current parameter package; for each recognition event, the day / night mode, available modal set, image quality, thermal imaging effectiveness, point cloud registration quality, positioning quality, time synchronization quality, vibration quality, trajectory continuity, fence segment binding quality, model version, and degradation reason are recorded.
[0136] In daytime mode, visible light cameras are the primary source of identification, while infrared or thermal imaging devices and lidar are secondary sources. In nighttime mode, passive vibrations from infrared or thermal imaging devices, lidar, and fence vibration sensors are the primary source of identification, while low-light visible light images are secondary sources.
[0137] In one embodiment, in night mode, the event level that has met the same evidence conditions is not reduced due to modality switching; when the complete identification conditions cannot be met only because some modalities are unavailable, the identification result and quality mark of the currently available modality are output; when the evidence is insufficient, the result to be verified is output.
[0138] In the above implementation process, by allocating the main identification source and auxiliary identification source according to the day and night mode, and recording the available modal set and various quality markers for each identification event, it is possible to output results with quality markers or results to be verified even when some modalities are unavailable or the evidence is insufficient at night, thus avoiding the substitution of deterministic conclusions with a single clock condition or low-quality modalities.
[0139] In one possible implementation, step S201 includes: acquiring multiple positioning results and determining the credibility of each positioning result; dynamically weighting and fusing the positioning results that have passed the consistency check, and outputting the current parameters of the perimeter inspection robot.
[0140] The reliability of the positioning results is configured to be updated based on the residuals across positioning sources.
[0141] The various positioning results here can include satellite positioning results, lidar positioning results, inertial measurement results, wheel speed odometer results, and visual positioning results, etc., and these multiple positioning results can be selected according to the actual situation.
[0142] In one embodiment of this application, the robot reads positioning results from BeiDou or GNSS, RTK, SLAM lidar, IMU, wheel odometer, and visual positioning, and maintains a reliability score for each positioning source. The reliability score is dynamically updated based on the number of satellites and differential status, point cloud matching residuals, number of visual features, IMU drift estimation, wheel-end slippage rate, and cross-positioning source residuals. Positioning sources that pass the consistency check are dynamically weighted or filtered and fused to output position, heading, velocity, attitude, and covariance.
[0143] By assigning credibility to the residual updates across positioning sources for satellite positioning results, lidar positioning results, inertial measurement results, wheel speed odometer results, and visual positioning results, and dynamically weighting and fusing the results that pass the consistency test, the positioning jump problem of a single positioning source under occlusion or degradation conditions is improved, which is conducive to the stable output of the pose and covariance required for boundary inspection.
[0144] Optionally, the absolute positioning weight is increased when the RTK fixed solution is valid and consistent with SLAM; its weight is decreased when satellite signals are blocked or abrupt, and continuous positioning is maintained by SLAM, vision, IMU, and wheel speed odometry; the wheel speed odometry weight is decreased when slippage occurs; and the visual weight is decreased when rain, fog, or low light cause visual degradation. The weight is gradually restored after the positioning source is recovered to avoid sudden position jumps.
[0145] By gradually restoring the positioning source weights instead of switching them immediately, body posture jitter caused by sudden changes in positioning results is avoided, which helps to maintain stability during touch loading.
[0146] The current parameters of the above-mentioned boundary inspection robot may include: the robot's position, heading, speed, attitude, and covariance, etc., which can be selected according to the actual situation.
[0147] In one specific embodiment of this application, the cruise controller projects the real-time position onto the target fence segment, calculates the lateral error of the robot from the inspection offset line, the heading error of the fuselage's major axis relative to the fence direction, and their confidence intervals, and simultaneously generates speed and steering control variables. Candidate paths intersect with the electronic fence; paths entering the runway, taxiway, or other restricted areas are rejected. The controller only allows entry into the touch control state when the uncertainty of the position and distance from the fence, attitude, and lateral space all meet the touch control permission conditions.
[0148] By intersecting candidate paths with the electronic fence and rejecting paths that enter restricted areas, the risk of accidentally entering the runway or taxiway during inspections is reduced; by treating uncertainty, attitude, and lateral space as touch permission conditions, the deployment of the touch module is avoided when safety conditions are not met.
[0149] In the above implementation process, by maintaining the reliability of each positioning source with dynamic updates of the residuals across positioning sources, and by dynamically weighting and fusing the positioning results that have passed the consistency test, the problem of the single positioning method being degraded under conditions of occlusion, slippage, rain, fog or low light affecting the continuity of perimeter inspection can be improved, which is conducive to the stable control of the distance and attitude between the robot and the fence.
[0150] In one possible implementation, the method further includes: controlling the perimeter inspection robot to decelerate and increase the safe distance between it and the fence when the confidence level drops to a first threshold; controlling the perimeter inspection robot to stop and reposition itself when the confidence level drops to a second threshold; and controlling the perimeter inspection robot to return along the confirmed trajectory or wait for remote takeover if the repositioning fails.
[0151] The second threshold is less than the first threshold. The first and second thresholds can be set according to actual circumstances.
[0152] It should be understood that by dividing the decline in credibility into two levels and corresponding to deceleration and increasing the safety distance and stopping in-situ repositioning respectively, the perimeter inspection robot can gradually tighten its movement permissions according to the positioning quality, reducing the risk of collision or invalid detection caused by continuing to approach the fence under uncertain positioning conditions.
[0153] Optionally, when there is no effective absolute reference for positioning but short-term relative positioning is still reliable, the perimeter inspection robot can return at low speed along the confirmed safe trajectory without entering new touchpoints; when the reliability reaches the stopping threshold, it maintains a stable posture and waits for repositioning or remote takeover. By restricting the perimeter inspection robot to only return at low speed along the confirmed trajectory without entering new touchpoints when there is no effective absolute reference, new detection records can be avoided in unreliable positioning states, which helps improve the reliability of inspection records.
[0154] In the above implementation process, by slowing down according to the level of confidence, increasing the safety distance, stopping repositioning and returning or waiting for remote takeover, the risk of the robot continuing to approach the fence when the positioning confidence is insufficient can be reduced, which will lead to collisions or false detections and improve the safety of the fence inspection process.
[0155] In one possible implementation, target recognition is performed on the image information to determine the target category of the fence intrusion target and to establish the target trajectory of the fence intrusion target. This includes: acquiring image information of the area inside and outside the fence through a visible light camera, infrared or thermal imaging equipment; using a detection and tracking model trained with airport scene samples to identify at least one target among people, dogs, birds and other animals in the image information; establishing a continuous trajectory for each identified target and outputting the target information.
[0156] The target information includes: target category, detection and tracking confidence level, location, speed, direction of movement, closest distance to the fence centerline, inner and outer markings, associated fence segment number, route coordinates, geographic coordinates, unified timestamp, and identification quality mark.
[0157] By establishing a continuous trajectory for the target and outputting the closest distance to the fence centerline, inner and outer identification, associated fence segment number, route coordinates, geographic coordinates, unified timestamp, and identification quality marker, image information is transformed into structured target information that can be associated with fence segment number, making it easier to participate in event association with touch diagnostic results under the same spatiotemporal index.
[0158] In the above implementation process, multimodal acquisition of the area inside and outside the fence is carried out by using visible light cameras and infrared or thermal imaging equipment. The detection and tracking model trained with airport scene samples is used to identify people, dogs, birds, other animals and airborne objects. A continuous trajectory is established for each target and the category, confidence level, position, speed, direction of movement, associated fence segment number, route coordinates, geographic coordinates, unified timestamp and identification quality mark are output. The image information is transformed into structured target information that can be associated with fence segment number and timestamp, which is conducive to providing evidence of intrusion targets under the same spatiotemporal index for subsequent cross-modal fusion.
[0159] In one possible implementation, the method further includes determining the climbing probability and crossing probability of the target trajectory based on the relationship between the target trajectory and the center line of the fence.
[0160] The detection results of fence intrusion targets include at least one of the following: target category, climbing probability, crossing probability, and vibration anomaly probability corresponding to the fence vibration signal.
[0161] The vibration signal of the fence is compared with the normal fence response benchmark to obtain the vibration anomaly probability. The detection results of fence intrusion targets include at least one of the following: target category, climbing probability, crossing probability, and vibration anomaly probability corresponding to the fence vibration signal.
[0162] By outputting detection results in the form of category plus probability, and specifying the necessary elements in a parallel manner with at least one, the detection results can still participate in event correlation even when some elements are missing, which helps to maintain the continuity of the alarm chain.
[0163] In the above implementation process, by establishing the target trajectory and determining the climbing probability and crossing probability based on its relationship with the center line of the fence, the visual detection results can participate in subsequent event association in the form of probability, which can improve the problem of only outputting the target category and not being conducive to event classification judgment.
[0164] In one possible implementation, the climbing probability and crossing probability of the target trajectory are determined based on the relationship between the target trajectory and the centerline of the fence, including: constructing a crossing determination surface with the centerline of the fence and the inner and outer directions; generating a concern event when the fence intrusion target is located outside the fence and moves parallel to it; increasing the concern level of the fence intrusion target when the fence intrusion target continues to approach the fence; forming a climbing probability when the outline of the fence intrusion target overlaps with the fence area and the height or position change of the target trajectory meets the climbing characteristics; and forming a crossing probability when the target trajectory continuously enters the inner side of the fence from the outer side and meets the spatial and duration conditions.
[0165] By constructing a crossing determination surface along the fence centerline and the inner and outer directions, and forming climbing and crossing probabilities based on contour overlap and height changes, as well as trajectory crossing and duration, the determination of intrusion behavior has verifiable spatial and temporal conditions.
[0166] In the above implementation process, by constructing a crossing determination surface with the fence centerline and the inner and outer directions, the motion state of the target relative to the fence is quantified into the event of concern, the level of concern, the climbing probability, and the crossing probability, respectively. This makes the determination of intrusion behavior verifiable in space and time conditions, which is conducive to converting the intrusion detection results into probabilistic evidence that can participate in quantitative fusion before cross-modal fusion.
[0167] In one possible implementation, under safe conditions, the perimeter inspection robot is controlled to enter a safe approach state and its body posture is adjusted, including: projecting real-time positioning coordinates onto the target fence segment to determine the lateral and heading errors of the perimeter inspection robot relative to the inspection offset line; determining the fence normal according to the fence direction associated with the fence segment number; under safe conditions, deploying the side touch detection module ③ and adjusting the body posture so that the side touch detection module ③ contacts the fence along the fence normal; under unsafe conditions, the perimeter inspection robot is controlled to reposition itself or collect preset data.
[0168] The safety conditions here include whether the location confidence, fence distance, fuselage attitude, support margin, and lateral space meet the touch permission requirements.
[0169] In one embodiment of this application, when the robot reaches the predetermined touch point, it adjusts its body posture according to the direction and lateral distance of the fence segment, switching the drive wheels to low-speed holding or parking mode. The touch module retracts along the body during normal cruising, such as... Figure 6 As shown; after entering the detection state, it extends towards the fence, as... Figure 7 As shown.
[0170] Optionally, the preset data may include visual data, thermal imaging data, and passive vibration data, which can be selected according to the actual situation.
[0171] Understandably, by projecting real-time positioning coordinates onto the target fence segment to obtain the lateral and heading errors relative to the inspection offset line, and then determining the fence normal using the fence direction associated with the fence segment number, the side touch detection module ③ can contact the fence along the fence normal, thus improving the force-displacement characteristic distortion caused by the loading direction deviating from the fence normal. Furthermore, by switching to passive acquisition instead of forced contact when safety conditions are not met, the fence segment can still retain visual, thermal imaging, and vibration evidence even when active touch conditions are unavailable, which helps maintain inspection coverage.
[0172] In the above implementation process, by projecting the real-time positioning coordinates onto the target fence segment and determining the contact direction with the fence normal, the touch loading direction is consistent with the force direction of the fence, which can improve the problem of incomparable detection waveforms caused by contact point offset or loading direction deviation; when the safety conditions are not met, it switches to passive acquisition, which is beneficial to retain usable evidence in fence segments where active touch is not possible.
[0173] In one possible implementation, step S203 includes: compensating for the displacement of the side touch detection module ③ and the base displacement based on the equivalent flexibility and contact force of the side touch detection module ③ to obtain an effective displacement.
[0174] The equivalent compliance was obtained by graded loading tests on rigid calibration components and grouped according to temperature or mechanism condition.
[0175] The base displacement here is obtained by joint estimation of the kinematics, inertial measurement attitude, and short-time positioning increment of the perimeter inspection robot.
[0176] In one embodiment, the controller synchronously acquires contact force, actuator-measured displacement, touch base displacement along the fence normal estimated by joint encoders and body pose, fence or probe vibration acceleration, fence inspection robot roll and pitch angles, foot force, and actuator current using a unified clock. Each channel records sampling time, range, saturation, missed samples, and synchronization error flags.
[0177] By synchronously acquiring contact force, actuator displacement, base displacement, vibration, attitude, foot force, and actuator current under a unified clock, and recording the sampling time and quality markers for each channel, traceable raw data is provided for subsequent displacement compensation and quality determination.
[0178] It should be understood that, to eliminate the effects of robot body movement and elastic deformation of the touch mechanism, the actuator displacement is converted into the effective displacement of the fence. The effective displacement can be calculated using the following formula: ; in, This represents the effective displacement of the fence along the loading direction. For the base position shift, This refers to the displacement of the touch base along the normal direction of the fence. For the equivalent flexibility of the touch mechanism, It is the contact force.
[0179] Understandably, by subtracting the base displacement and the elastic deformation of the mechanism calculated according to the equivalent flexibility and contact force from the actuator displacement, the measured displacement is converted into the effective displacement of the fence itself, so that the equivalent stiffness, hysteresis and residual displacement have a basis for comparison across actions and across equipment; by grouping the equivalent flexibility according to temperature or mechanism state, the influence of mechanism state changes on displacement compensation can be reduced.
[0180] In the above implementation process, by compensating for the elastic deformation of the touch mechanism and the motion of the robot base respectively, the displacement measured by the actuator is converted into the effective displacement of the fence, so that the force-displacement characteristics obtained by different actions and different devices have a comparable basis, which helps to reduce the influence of the robot's own motion mixed into the detection results.
[0181] In one possible implementation, determining the fence state characteristics based on the effective displacement includes: determining the equivalent stiffness and normalized hysteresis based on the contact force and the effective displacement, and determining the residual displacement after unloading; performing time-frequency decomposition on the fence vibration signal after removing the base movement to determine the energy ratio and vibration attenuation coefficient of each frequency band; and determining the fence state characteristics based on at least one of the equivalent stiffness, normalized hysteresis, residual displacement, energy ratio, vibration attenuation coefficient, and spatial consistency of adjacent detection points.
[0182] In one embodiment, the equivalent stiffness can be obtained by the following formula: ; in, For equivalent stiffness, To select the contact force within the loading range, This represents the effective displacement of the fence.
[0183] Here, the fitting residuals, effective sample size, and coefficient of determination are used as quality indicators. Notably, if the fitting quality is unsatisfactory, no deterministic stiffness conclusion is output.
[0184] It should be understood that by performing robust linear fitting after excluding contact establishment and saturation intervals, and using the fitting residuals, effective sample size, and coefficient of determination as quality markers, deterministic stiffness conclusions can be avoided when the fitting quality is insufficient.
[0185] Normalized hysteresis can be calculated using the following formula: ; in, To normalize the hysteresis, the integral is calculated along the load-unload closed path. This represents the maximum contact force during the action. For the maximum effective displacement, To prevent calibration of small quantities with a denominator of zero.
[0186] The normalized hysteresis here is used to reflect connection friction, relaxation and irreversible energy loss, and is used in conjunction with the residual displacement after unloading.
[0187] It should be understood that by integrating the loading-unloading closed path and normalizing it with the maximum contact force and the maximum effective displacement, the hysteresis characteristics under different loading amplitudes can be compared, which is beneficial for reflecting the changes in connection friction, relaxation and irreversible energy dissipation.
[0188] Furthermore, wavelet packet decomposition is performed on the vibration signal after removing the low-frequency base motion, and the energy and energy ratio are calculated in each frequency band. The energy ratio can be expressed by the following formula: ; in, For the first The first wavelet packet frequency band One coefficient, For the energy of this frequency band, This represents the normalized energy ratio.
[0189] The wavelet basis, number of decomposition layers, and frequency band boundaries here are selected from calibration samples of normal mesh, loose, broken wires or detached welds, deformed, and abnormal posts, avoiding the pre-setting of arbitrary frequency bands. By determining the wavelet basis, number of decomposition layers, and frequency band boundaries using calibration samples, rather than pre-setting fixed frequency bands, the frequency band energy ratio can correspond to the response differences of actual fence structures, which is beneficial to improving the versatility between different fence types.
[0190] The vibration envelope after impact was fitted with an exponential decay model: ; in, The amplitude is obtained through the Hilbert envelope or equivalent envelope method. The initial amplitude, This represents the attenuation coefficient. A smaller attenuation coefficient indicates slower attenuation.
[0191] It should be understood that by performing exponential decay fitting on the vibration envelope and setting calibration conditions for participation in diagnosis, the decay coefficient can be avoided from being used for defect determination when the impact energy is insufficient or the fitting quality is poor.
[0192] Optionally, the fence state characteristics consist of at least one of the following: equivalent stiffness, normalized hysteresis, residual displacement, energy ratio of each frequency band, vibration attenuation coefficient, and spatial consistency of adjacent detection points.
[0193] Among them, the residual displacement is the offset of the effective displacement of the fence relative to the zero force position after unloading, and the spatial consistency of adjacent detection points is the consistency of the characteristic change direction of adjacent detection points within the same fence segment.
[0194] Understandably, by combining static force-displacement characteristics with dynamic vibration characteristics and spatial consistency characteristics into fence state characteristics, the physical state of the fence can be described with multi-dimensional features, providing a basis for subsequent discrimination based on defect categories.
[0195] In the above implementation process, the fence state characteristic quantity is formed by combining the force-displacement curve characteristics, unloading residue, vibration time-frequency energy and attenuation characteristics, and spatial consistency of adjacent detection points. The empirical inspection of manually shaking the fence is transformed into repeatable and calibrable quantitative characteristics, which helps to improve the problem that a single contact force threshold is difficult to distinguish the defect type.
[0196] In one possible implementation, the method further includes: determining a baseline group according to the fence parameters corresponding to the fence segment number; standardizing the fence state feature quantity using the median and median absolute deviation of the baseline group; returning to the previous level baseline group if the baseline samples are insufficient; and stopping the output of fence defect categories if the data quality is unacceptable.
[0197] The fence parameters can include fence structure type, mesh material, post spacing, temperature range, and wind conditions. These parameters can be selected based on actual conditions.
[0198] Optionally, if the data quality is substandard, a data quality substandard flag can also be set.
[0199] In one embodiment, the standardization of the fence state characteristics can be achieved by the following formula: ; in, Indicates the benchmark group of the match. For the benchmark group The median of the features, This represents the median absolute deviation of the baseline group.
[0200] Optionally, if the baseline sample is insufficient, it is possible to fall back to the next higher baseline of the same type of fence, but the confidence level of the conclusion must be reduced and the baseline level used must be recorded.
[0201] It should be understood that robust standardization using the median and median absolute deviation can reduce the impact of individual outliers on the baseline; and it can revert to the next higher baseline group when the baseline sample is insufficient. Simultaneously, it can lower the diagnostic confidence level, record the baseline level used, and ensure that the reliability of the conclusion corresponds to the granularity of the baseline on which it is based, facilitating subsequent review and updates.
[0202] The aforementioned data quality deficiencies may include: critical channel saturation, excessive synchronization error, reference mismatch, contact point slippage, or insufficient action repeatability. The appropriate data quality deficiencies can be selected based on the specific circumstances.
[0203] Optionally, if any critical channel is saturated, synchronization error exceeds the limit, reference mismatch occurs, contact point slippage occurs, or action repeatability is insufficient, the system marks the data quality as unqualified and stops outputting deterministic categories; if the contact force, mechanism stroke, machine roll or pitch, support margin, actuator current, or fence distance consistency exceeds the safety threshold, loading, unloading, and retraction are stopped immediately.
[0204] Understandably, by directly linking data quality deficiencies to diagnostic outputs, deterministic defect categories can be avoided when no valid data is available; by associating safety constraint breaches with immediate unloading and retraction, the risk of damage to the fence and robot during the loading process is reduced.
[0205] In the above implementation process, benchmarks are established by grouping according to structural type, material, column spacing, temperature range and wind conditions, and robust standardization is performed using median and median absolute deviation to reduce the interference of abnormal samples on the benchmarks. When the benchmark samples are insufficient, the data is returned to the previous level benchmark group and the diagnostic confidence is reduced and the benchmark level is recorded. When the data quality is unqualified, the output of the defect category is stopped, which helps to avoid outputting unfounded definitive conclusions.
[0206] In one possible implementation, step S204 includes: determining the fence defect category as mesh loosening when at least one of the following is true: the equivalent stiffness is lower than the normal fence response benchmark, the normalized hysteresis is increased, the vibration attenuation coefficient is decreased, or the energy ratio exceeds the normal frequency band, and the anomaly occurs at the mesh detection point.
[0207] If a sudden change occurs during the loading process, and it is consistent with the sudden change at the same location in the preset state, the fence defect category is determined to be broken wire or detached weld.
[0208] If the residual displacement exceeds the residual displacement threshold and the zero-force position of repeated touch actions remains in the same direction of deviation, the fence defect category is determined to be mesh deformation.
[0209] If low-frequency displacements in the same direction occur on both sides of the same post or at multiple adjacent detection points, the compensated base displacement is insufficient to explain the low-frequency displacements, and the low-frequency energy or attenuation characteristics exceed the normal reference of the post, the fence defect category is determined to be post instability.
[0210] Optionally, the touch action can be a quasi-static push, a reciprocating micro-motion, or a pulse contact of a specified number of times, and the touch action can be selected according to the actual situation.
[0211] In one embodiment, the comprehensive anomaly score is calculated from calibrated linear terms and necessary interaction terms: ; Among them, σ(u)=1 / (1+e^( u)), This is a comprehensive score based on the probability of touch anomalies or after probability calibration. , and This is to fit the data to the calibration sample or training set.
[0212] The aforementioned defect categories can be trained separately for category scores, or they can be output by the classification model after rule gating.
[0213] It should be understood that by inputting the standardized features and their interaction terms into the calibrated scoring function, the comprehensive anomaly score is output in probabilistic form, which facilitates its participation in subsequent fusion on the same dimension as the confidence scores of visual and vibration modes.
[0214] The residual displacement threshold here can be generated by the normal baseline quantile, the subject working characteristics of the labeled defective samples, or a cost function that meets airport safety requirements.
[0215] Optionally, the residual displacement threshold can be stored along with the fence type, environment range, and model version.
[0216] The preset states include: visual linear gaps, missing connection points, or adjacent repetitive actions. These preset states can be selected according to the actual situation.
[0217] Abrupt changes during the loading process described above can include: sudden drops in local force, sudden changes in slope, or increases in high-frequency transient energy during loading or impulse response. The abrupt change pattern during the loading process can be selected based on the actual situation.
[0218] It should be understood that by setting criteria composed of force-displacement characteristics, vibration time-frequency characteristics, and spatial distribution characteristics for the four types of defects, and by using spatial gating to distinguish between mesh loosening and column instability, the problem of difficulty in distinguishing between local mesh loosening and low-frequency swaying of the support structure based solely on the contact force threshold is improved. This makes the defect category judgment repeatable, calibrable, and locatable.
[0219] By binding and storing thresholds with fence type, environmental range, and model version, the same criterion can be calibrated and updated separately under different airports, different fence structures, and different environmental conditions, thus avoiding treating uncalibrated parameters as fixed product performance.
[0220] In the above implementation process, by setting executable criteria for mesh loosening, wire breakage or detachment, mesh deformation and column instability, and by using spatial gating of "local mesh anomaly" and "consistent anomaly across adjacent detection points" to distinguish between mesh loosening and column instability, the problem that local loosening and low-frequency oscillation of the support structure cannot be distinguished by contact force threshold alone can be improved, which is conducive to improving the distinguishability of defect category judgment.
[0221] In one possible implementation, step S205 includes: determining the identification results of personnel and biological intrusion, the physical inspection results corresponding to the fence defect category, and the fence vibration anomaly results as an event association set, based on the fence segment number being the same or belonging to adjacent fence segments that are allowed by the calibration and whose time difference is within the time window; classifying the confirmed events into personnel or biological intrusion events, personnel climbing events, high-intensity impact events, fence physical state abnormality events, and passage risk events based on the confirmed event type; and generating observation, attention, warning, or emergency alarms based on the confirmed event type.
[0222] In one embodiment, before event association, the system projects the trajectory of each visual target onto a fence segment and route coordinates, establishing an event window centered on the time the target enters the fence neighborhood. Passive vibration, active touch, and fixed perimeter sensor results are associated with the target only if the fence segment is the same or belongs to a calibrated adjacent segment, the time difference is within the window, and their respective quality markers are valid. In case of association conflicts, the system selects the highest association probability based on spatial distance, time difference, and trajectory continuity, avoiding the repeated assignment of the same vibration evidence to multiple irrelevant events.
[0223] By limiting the association to only those fence segments that are the same or belong to adjacent fence segments that are allowed by the calibration, and whose time difference is within the time window, and by selecting the maximum association probability based on spatial distance, time difference and trajectory continuity when there is an association conflict, it is possible to avoid the same vibration evidence being repeatedly assigned to multiple unrelated events, which is beneficial to suppressing false alarms caused by common cause vibrations such as wind blowing vegetation.
[0224] In one embodiment, the alarm level is determined by the event type and confirmed evidence: data that does not form a valid event association is only recorded as observations; when an event of concern is formed, the subsequent sampling frequency is increased; confirmed intrusion, physical status, or security risk events generate warnings requiring manual review and upload evidence summaries; events that meet the conditions for an emergency alarm generate an emergency alarm and are transmitted with priority.
[0225] By classifying events into observation, attention, warning, or emergency alerts based on event type and confirmed evidence, different handling and transmission strategies can be applied to events with different confidence levels. This helps reduce the handling costs of low-confidence events without concealing urgent events.
[0226] In the above implementation process, by establishing event associations according to fence segment numbers and time windows, and by unifying the identification results of personnel and biological intrusions, the physical inspection results corresponding to fence defect categories, and the fence vibration anomaly results into an event association set, and outputting observation, attention, warning, or emergency alarms according to the confirmed intrusion, physical state, or security risk event types, the problem of inconsistent spatiotemporal references of data from different sensors and the difficulty in cross-verifying can be improved. By classifying by event type instead of relying on a single confidence threshold, it is beneficial to suppress false alarms caused by a single visual or single vibration noise without masking emergency events.
[0227] In one possible implementation, the method further includes: if any sub-module in personnel and biological intrusion identification, fence defect category diagnosis, or fence vibration detection is unavailable or has substandard identification quality, the identification result of the available modality or available sub-module is downgraded to participate in event correlation, and the reason for the missing, the reason for the downgrade, and the quality mark are recorded; if safety conditions do not allow active touch control, the alarm level is determined solely based on visual, thermal imaging, and passive vibration evidence.
[0228] An emergency alarm is generated directly if the target trajectory remains inside the fence, or if the target trajectory moves from outside the fence to inside the fence and the duration meets the safe duration requirement.
[0229] In one embodiment, when a certain modality is missing or of substandard quality, if any submodule in personnel and biological intrusion identification, fence defect category diagnosis, or fence vibration detection is unavailable or of substandard quality, the identification result of the available modality or available submodule is downgraded to participate in event association, and the reason for the missing modality, the reason for the downgrade, and the quality mark are recorded; if the safety conditions do not allow active touch control, the alarm level is determined only based on visual, thermal imaging, and passive vibration evidence; if the target trajectory is continuously located inside the fence, or if the target trajectory enters the fence from the outside and the duration meets the safety duration condition, an emergency alarm is directly generated.
[0230] By using an available modality or an available submodule in event correlation when an identification submodule is unavailable or its identification quality is substandard, and recording the reasons for the missing module, the reasons for the demotion, and the quality flags, the distortion of conclusions caused by using the complete model after setting the missing probability to zero can be avoided. By determining the alarm level solely based on visual, thermal imaging, and passive vibration evidence when active touch is not permitted under safe conditions, alarm judgment is not dependent on the prerequisite of necessary contact.
[0231] Optionally, confirming that the target trajectory enters the inner side from the outside and continues for more than the safe time, that the target remains inside the perimeter, that the fixed perimeter system issues a certified high-level alarm, or that the fence experiences a high-intensity impact exceeding the structural safety threshold can serve as high-risk independent evidence to directly trigger an emergency alarm, unaffected by the temporary absence of other modalities or the dilution of fusion probability.
[0232] By directly generating an emergency alarm when the target trajectory remains inside the fence or enters the fence from outside and the duration meets the safe duration condition, high-risk events are not affected by the temporary absence of other modalities, which helps to shorten the reporting delay of emergency events.
[0233] Optionally, the alarm evidence package includes: event number, target trajectory identifier, category, fusion confidence level, independent trigger marker, occurrence time, fence segment number, route coordinates, geographic coordinates, key images, video index before and after the event, vibration and touch feature summary, input quality, model version, and robot status, etc.
[0234] By including fusion confidence, independent trigger markers, input quality, and model version in the alarm evidence package, the platform can determine the credibility of the alarm conclusion and arrange for review, which is beneficial for the traceability and auditing of inspection records.
[0235] In the above implementation process, by using available modalities or available sub-modules to participate in event association and recording the reasons for the missing, the reasons for the downgrade, and the quality markers when the identification sub-module is missing or the identification quality is unqualified, the distortion of conclusions caused by using the complete model after setting the missing probability to zero can be avoided; by setting high-risk independent evidence to directly trigger emergency alarms, it is beneficial to be able to report emergency events in a timely manner when some modalities are temporarily missing.
[0236] In one possible implementation, the perimeter inspection robot is a multi-legged wheeled robot, which includes a body, multiple mechanical legs, and a drive wheel at the end of each mechanical leg. The method also includes: establishing a local elevation map based on laser and visual signals, and determining a terrain score of the terrain where the multi-legged wheeled robot is located by combining the current motion parameters and terrain parameters of the multi-legged wheeled robot; and determining the motion mode of the multi-legged wheeled robot based on the terrain score.
[0237] The sports modes include: wheeled high-speed driving mode, low-speed wheeled driving mode, wheel-foot coordinated obstacle crossing mode, and safety stop mode.
[0238] In one embodiment, the robot travels rapidly on level terrain using wheels, with each drive wheel operating at the target linear velocity and steering angle, while the robotic legs maintain a basic posture that balances energy consumption, ground clearance, and sensor field of view. Before entering grass, mud, potholes, ditches, slopes, or obstacle areas, the controller uses laser point clouds and visual depth to create a local elevation map, and calculates a passability score by combining the robot's roll angle, pitch angle, wheel speed, foot force, and slip rate.
[0239] By establishing a local elevation map based on laser and visual signals before entering uneven areas, and combining the current motion parameters and terrain parameters to determine the terrain score, the selection of motion modes is based on quantifiable terrain descriptions rather than relying on manual remote control judgment.
[0240] Each input, after time alignment, outlier removal, and normalization, forms the slope, roughness, ditch width and depth, wheel-to-ground adhesion, fuselage stability margin, and landing feasibility factor. The state machine with hysteresis switches between high-speed wheeled travel, low-speed wheeled travel, wheel-foot coordinated obstacle crossing, and safe stopping based on the score and safety constraints, avoiding frequent oscillations near the threshold.
[0241] By using a state machine with hysteresis to switch between four motion modes, frequent mode switching caused by score fluctuations around the threshold can be avoided, which helps to maintain the stability of the aircraft's attitude and the sensor's field of view.
[0242] In one specific embodiment of this application: when the passability score is higher than a first calibration threshold and the slippage rate is lower than the upper limit, the wheeled vehicle maintains a fast speed according to the fence route speed limit. When the roughness, slope, or slippage rate increases but continuous rolling conditions are still available, the wheel speed is reduced, and the four leg joints are adjusted to maintain the body attitude and center of gravity projection. When ditches, steps, or insufficient continuous rolling conditions are detected, the vehicle switches to wheel-leg coordinated obstacle crossing, plans the wheel end landing point and lifting trajectory, and locks or controls some drive wheels at low speed if necessary. When the stability margin, remaining torque, landing feasibility, or sensor reliability are lower than the safety threshold, the vehicle stops moving forward and replans; if it still cannot pass, it detours or returns along a verified route.
[0243] By setting entry conditions for each of the four movement modes and linking them with safety thresholds, the robot can autonomously select between a high-speed wheeled driving mode, a low-speed wheeled driving mode, a wheel-foot cooperative obstacle-crossing mode, or a safe stopping mode based on the terrain score. This improves the problem that a single movement mode cannot simultaneously achieve both long-distance efficiency and the ability to traverse complex terrain.
[0244] Please see Figure 5 , Figure 5This is a schematic diagram of the wheel-foot composite motion and terrain adaptation process provided in the embodiments of this application. When the robot continuously travels between flat ground, local ditches, and undulating terrain, it maintains stability through changes in leg joint posture and drive wheel control; the dashed posture indicates the position before and after mode switching or the intermediate position in gait planning.
[0245] In the above implementation process, by determining the terrain score based on the local elevation map and the current motion parameters and terrain parameters, and switching the motion mode between wheeled fast driving, low-speed wheeled driving, wheel-leg coordinated obstacle crossing and safe stopping according to the terrain score, the problem of insufficient obstacle crossing ability of wheeled robots in grass, mud and ditch sections and the problem of limited speed and endurance of legged robots on long-distance flat roads can be improved, which is conducive to maintaining continuous long-distance boundary patrol.
[0246] In one possible implementation, the airport perimeter segment map includes multiple fence segments.
[0247] Each fence segment here has a unique number, and each fence segment is associated with fence information and area information.
[0248] The fence segment number, route coordinates, location coordinates, and unified timestamp are stored together.
[0249] Optionally, the fence information includes: fence type, mesh and post locations, inner and outer directions, inspection offset lines, etc. The area information includes: permitted access areas, prohibited areas, key detection points, normal fence response benchmarks, visual calibration parameters, communication coverage segments, and available cell sets, etc. This fence information and area information can be selected according to actual needs.
[0250] By associating the mesh and post positions, inner and outer directions, and inspection offset lines on each fence segment, a geometric basis is provided for determining the normal direction and calculating the lateral error during the safe approach phase; by associating the normal fence response benchmark and visual calibration parameters, a segment-based configuration benchmark is provided for diagnosis and identification; by associating the communication coverage segments and the set of available cells, a segmented candidate cell range is provided for communication prediction and pre-scheduling.
[0251] By associating and storing fence segment numbers, route coordinates, and location coordinates with a unified timestamp, positioning, touch, vision, vibration, alarm, and communication data can be retrieved and compared using the same spatiotemporal index, which is beneficial for post-event review, defect trend analysis, and model updates.
[0252] In the above implementation process, by configuring a unique number for each fence segment and associating it with geometric, benchmark, calibration and communication attributes, positioning, touch, vision, vibration, alarm and communication data are bound to the same spatiotemporal index, which can improve the problem of data from different sources being difficult to correlate with each other, and is conducive to the traceability and verification of inspection records.
[0253] In one possible implementation, the method further includes: during task initialization, loading the normal appearance reference, mesh geometric reference, connection point reference, sign position and attitude reference, perimeter box external closure reference, pole and box external geometric reference, foundation neighborhood reference, normal fence response reference, and applicable parameter version according to the fence segment number; projecting the current positioning coordinates to the target fence segment or facility route coordinates, and forming an observation pose consistent with the fence normal, inspection offset line, and safety distance; and adjusting the position confidence, lateral distance uncertainty, fuselage attitude, lateral space, support margin, or target safety distance. If the requirements are not met, the system maintains a non-contact detection mode using vision, thermal imaging, lidar, and passive vibration. It collects visible light images, infrared or thermal images, laser point clouds, inertial measurement attitude data, fence vibration signals, and operational status from multiple observation poses for the same fence section or facility. The laser point cloud is then registered with the corresponding geometric reference, and the registration result and registration quality marker are output. If candidate anomalies appear in vision, thermal imaging, laser point cloud, or passive vibration, and safety conditions are met and there is no risk of personnel, animals, or unknown targets near the fence, the side touch detection module is invoked to verify the safe-to-access detection points.
[0254] The inspection results of the fence and its ancillary facilities are bound to the task number, event category, fence segment number, route coordinates, geographical coordinates, unified timestamp, sensor source, input quality mark, benchmark version, parameter package version, handling status, and record status.
[0255] Alternatively, fence accessories may include poles, boxes, low-lying areas, etc., and these accessories can be selected according to the actual situation.
[0256] The inspection items for the aforementioned fence body include fence damage inspection, fence corrosion, fence screw loss, foreign objects hanging on the fence, stability of signs and other appearance abnormalities, and at least one of the following: fence perimeter boxes. The inspection items for intermediate poles and facilities include the integrity of monitoring poles, the integrity of light poles, intermediate power distribution boxes, and power distribution boxes within the monitoring station.
[0257] In the above implementation process, by binding the inspection results of the fence and its ancillary facilities to the fence section number, benchmark version, parameter package version and input quality mark, and limiting active touch control to a verification method only when safety conditions are met and there are no people, animals or unknown targets near the fence, structural condition evidence can be obtained without increasing the risk of contact. This is conducive to making the external condition conclusions of the fence body, poles, facilities and foundation traceable and verifiable.
[0258] In one embodiment, during the task initialization state, the external geometric reference, visible appearance reference, basic neighborhood reference, allowed observation pose and applicable parameter package corresponding to the monitoring pole, light pole, intermediate power distribution box and power distribution box in the monitoring station are loaded; during the boundary positioning state and safe approach state, the location is located to the corresponding fence segment or route coordinates of the facility, and multiple observation poses are obtained without entering the restricted area.
[0259] In one possible implementation, the fence damage inspection is triggered when the visible light image shows abnormal grid continuity, linear gaps or disconnection features at the connection points, the laser point cloud shows fence outline shift, local missing or abnormal openings, or the passive vibration shows abnormal impact or continuous vibration. The image features are associated with the point cloud features, and the output is pending verification when there is only a single mode and occlusion, shadows, vegetation, reflection or missing point cloud cannot be ruled out. Fence corrosion is triggered when the current visible light texture, color, surface mottled features, or changes in the appearance of the connection parts are abnormal relative to the normal appearance benchmark of the same fence type and the same environmental range. The touch results are only used as auxiliary evidence of the structural state and should not replace the judgment of the appearance of corrosion. The inspection of fallen fence screws is triggered when the visible appearance of the connection point is inconsistent with the reference of the connector, the relative position of adjacent components changes, or the point cloud has local geometric anomalies. The inspection is also triggered by equivalent stiffness, normalized hysteresis, residual displacement and vibration characteristics to distinguish between fallen connectors, loose connectors and general obstruction. Foreign object suspension fence inspection: When the visible object still maintains spatial association with the fence mesh or signage area under multiple observation poses, a foreign object suspension event is formed. When the target is in a continuous moving state and the target trajectory crosses the fence centerline judgment surface from the outside to the inside, it is transferred to the airborne object intrusion identification. The perimeter inspection of the enclosure only identifies the external outline of the enclosure, the door cover plane, the closed boundary, the tilt status, the connection relationship of the surrounding fence, the basic neighborhood, and visible external damage. It does not open the enclosure, touch the electrical components inside the enclosure, or read the internal electrical status.
[0260] In one possible implementation, the complete inspection of monitoring poles is used to identify the absence, tilting, sagging, or obvious detachment of poles, equipment suspension positions, speakers, supplementary lights, audible and visual alarms, lightning rods, and their visible external installation status; the complete inspection of light poles is used to identify abnormalities in the centerline of the light pole, the position of the light fixture, the posture of the light fixture, the sagging status of the light fixture, the tilt of the pole, and the relative position of the foundation, and does not interpret the absence of thermal response as a failure or absence of the light fixture; the inspection of intermediate distribution boxes and distribution boxes in monitoring stations is only used to identify the outline of the external box, the closed status of the door, the tilt of the box, the water accumulation in the vicinity of the foundation, and the relative position of the foundation, without opening the box, touching the internal wiring, reading the electrical status inside the box, and not presuming the external temperature map of the box, the appearance of the door, or the condition of the foundation as an internal power distribution fault; in cases where equipment is obstructed, point cloud registration is insufficient, the quality of nighttime images is unqualified, or the positioning uncertainty affects the facility's attribution, the status of the corresponding facility is output for verification.
[0261] In one possible implementation, the method further includes: identifying visible water accumulation candidates around low-lying areas, box foundations, and pole foundations; recording quality defects where water surface reflection, absorption, occlusion, or sparse point clouds lead to unreliable laser results, without interpreting missing point clouds as water accumulation; forming water accumulation candidates when multi-view visual or thermal imaging evidence is consistent with the low-lying area baseline and the point cloud or ground geometry supports the corresponding positional relationship; otherwise, outputting the low-lying area status for verification; identifying continuous geometric offsets of the edges, adjacent columns, mesh baselines, and ground elevation relationships of fence foundations, pole foundations, and box foundations relative to historical baselines; forming foundation settlement candidates when continuous and unidirectional geometric offsets are present and registration quality, positioning quality, and baseline matching conditions meet requirements; and collecting long-distance evidence while maintaining a safe distance when local ground accessibility requirements are not met.
[0262] Among them, the effective displacement, equivalent stiffness, normalized hysteresis, vibration energy, vibration attenuation coefficient, residual displacement and spatial consistency obtained by active touch are only used as auxiliary structural evidence.
[0263] In one possible implementation, the steep road section passability test function is applicable to anti-slip test sections deployed on the side of slopes. The task sheet should clearly specify the 25kg specification: it can be the minimum load-bearing capacity of the test section structure or the total mass of the robot in the test state; the two should not be used interchangeably. If 25kg represents the load-bearing capacity of the section structure, the relationship between the minimum load-bearing capacity and the total mass should also be recorded; if 25kg represents the total mass of the robot, the load-bearing capacity of the section structure should be confirmed separately. The section width and load-bearing conditions are recorded as parameters for the acceptance scenario deployment and do not constitute a fixed performance limitation for the robot body, mechanical legs, drive wheels, or control modules.
[0264] In the task initialization state, the robot controller ② loads the route coordinate range, permissible boundaries, test direction, route benchmark, passable scoring parameter package, and safe stop strategy of the test section, and writes the actual road width, slope range, effective length, surface condition, load-bearing diameter, total test mass of the robot, environmental conditions, test direction, parameter version, and expected passage method into the test record; in the boundary positioning state, it locates the entrance of the test section, and in the safe approach state, it confirms the body attitude, lateral safety space, positioning reliability, and return path; if the route coordinates, boundary position, or positioning reliability does not meet the requirements, it does not enter the road section, only records "steep road section to be verified" and executes safe stop or return.
[0265] The robot controller ② uses LiDAR to create a local elevation map in touch-controlled or passive detection mode. Combined with data from visible light, infrared or thermal imaging, IMU, wheel speed sensors, and foot contact force sensors, it obtains input evidence regarding slope, roughness, ditch width and depth, wheel-ground adhesion status, vehicle stability margin, footing feasibility, and slippage rate. Visible light is used to assist in identifying surface textures, boundaries, and obstacles; at night, LiDAR, IMU, wheel speed, foot contact force, and thermal imaging are the primary methods, with low-light visible light images serving only as supplementary data. For ground attributes that cannot be reliably estimated using existing sensor quality, the robot controller ② reduces the reliability of the drivability score and does not assign it a fixed empirical value.
[0266] The robot controller ② makes a comprehensive judgment based on terrain drivability scores and safety constraint logic during state diagnosis. Entry thresholds, exit thresholds, factor weights, and hysteresis intervals are all generated from normal road sections, calibration samples, or training sets, and uncalibrated fixed values are not used. When the drivability score meets the conditions for high-speed wheeled travel, and the slip rate, attitude, and boundary constraints meet the requirements, wheeled travel is maintained. When the slope, roughness, or slip rate increases but the continuous rolling condition is still met, low-speed wheeled travel is switched. When ditches, steps, insufficient continuous rolling, or more reliable landing plans are encountered, wheel-foot cooperative mode is switched. When the wheel-ground adhesion status, body stability margin, landing feasibility, positioning quality, lateral boundary, or sensor quality do not meet safety conditions, a safe stop is initiated. The switching employs hysteresis rules to prevent frequent switching between wheeled, wheel-foot cooperative, and stop modes when the score is within a critical range.
[0267] In M5 state, robot controller ② associates terrain point cloud, attitude change, wheel speed change, foot contact force, slippage state, mode switching record, and fence segment number; in M6 state, it generates operation status record for situations that are passable but require low speed or wheel-foot coordination, and generates warnings or emergency safety handling events for situations that are unsafe to pass, repeated slippage, boundary risks, or insufficient stability margin; in M7 state, it records passage score, mode switching, and handling summary; in M8 state, it executes safe stop, back to verified route, detour, return, or remote takeover.
[0268] The acceptance criteria for this test function are as follows: The robot controller ② should completely record the fence segment number or station route coordinates, geographic coordinates, unified timestamp, terrain quality markers, passage score results, wheeled or wheel-foot cooperative mode, slippage status, reason for stopping or resuming, parameter version, and actual road width, slope range, effective length, and load-bearing diameter at the entrance, road segment, and exit. Only when the robot completes the prescribed actions of entering, continuous passage, controlled deceleration or braking, and leaving the road segment within the test direction and route boundaries specified in the task sheet, and without triggering stability margin, lateral boundary, continuous slippage, critical sensor failure, or safety control boundary violations, will it be recorded as "passed". When road conditions exceed the calibration range or sensor quality is insufficient, a safe stop or "no valid conclusion" should be selected; a safe stop, detour, or failure to pass should not be interpreted as safe passage.
[0269] In the above implementation process, by writing the road width, slope, load-bearing diameter and total test mass into the test record and clearly stating that "failure to pass must not be interpreted as safe passage", it is possible to avoid misinterpreting safe stopping or detour as passing the test, which helps to make the passing test conclusion correspond to the actual road conditions.
[0270] In one possible implementation, the method further includes: establishing scenario-based functional verification records according to the scenario, time pattern, detection object, expected output and handling status; in the case of lack of effective normal baseline, current environment being an out-of-sample environment, sensor quality not meeting the conditions, parameter package expired, model version mismatch, or event location not being reliably bound to the fence segment, outputting no valid conclusion or pending verification, and saving available original evidence.
[0271] Each record must be associated with at least the task number, fence segment or site route coordinates, geographic coordinates, unified timestamp, parameter package version, sensor availability, original evidence index, identification quality marker, conclusion status, alarm level, record status, and handling result. The conclusion status is categorized as "Confirmed," "Candidate for Verification," "No Valid Conclusion," and "Safety Aborted." "Confirmed" requires that currently available evidence meets the quality, relevance, and repeatability conditions specified in the scenario's parameter package. "Candidate for Verification" indicates that there are abnormal clues but are insufficient to rule out alternative explanations such as occlusion, insufficient viewpoint, nighttime imaging degradation, sparse point cloud, baseline mismatch, or incomplete work permit information. "No Valid Conclusion" indicates that key data or baselines are unavailable. "Safety Aborted" indicates that the robot did not execute or interrupted related actions due to posture, terrain, target safety distance, or risks to key sensors. Any status other than "Confirmed" must not be described as a completed normal or abnormal confirmation in the task report.
[0272] In addition to fence baselines and model weights, the parameter package also includes scene-matched observation pose, effective observation distance, minimum identifiable target size, image sharpness, exposure quality, point cloud sampling density, visibility ratio, target occlusion ratio, number of repeated observations, trajectory continuity conditions, positioning and time synchronization quality, event association window, alarm response time limit, and evidence retention strategy.
[0273] In one embodiment, when there is no valid absolute reference for positioning but short-term relative positioning is still reliable, the robot can return at low speed along the confirmed safe trajectory without entering a new touch point; when the positioning confidence is lower than the stopping threshold, it maintains a stable posture and waits for repositioning or remote takeover. When the touch reference is missing, the sensor is saturated, the synchronization error exceeds the limit, or the safety unloading is triggered, the diagnostic record is marked as "no valid conclusion" and the reason and available original evidence are saved; the system can arrange manual re-inspection or re-inspection after the conditions are restored, and will not automatically interpret the lack of conclusion as the fence being normal.
[0274] When vision is affected by rain, fog, backlight, obstruction, or nighttime imaging degradation, the system lowers the quality label of the corresponding recognition result, and can use available evidence such as thermal imaging, lidar, and passive vibration. When personnel, animals, or unidentified targets are near the fence, approaching the target to obtain active touch evidence is prohibited, and safe distance observation and alarms are prioritized. If the risk to positioning, attitude, power, or critical sensors continues to exceed the policy, the system will enter a safe stop, return to base, or remote takeover.
[0275] For scenarios involving external facilities and personnel, the system will only make conclusions regarding the external physical condition or authorized personnel verification within the scope of the task, parameter packages, and data quality. The electrical health of the enclosure, personal identification information, and business status beyond the scope of the fence inspection task will not be presumed based on vision, thermal imaging, point clouds, vibration, or a single external signal.
[0276] In the above implementation process, by distinguishing the conclusion status into confirmed, candidate pending verification, no valid conclusion and safety suspension, and stipulating that non-confirmed status cannot be described as completed confirmation, it is possible to avoid describing the result of insufficient evidence as the fence being normal, which is conducive to the consistency between the inspection report and the actual evidence status on site.
[0277] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.
[0278] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.
[0279] If a function is implemented as a software module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk. It should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0280] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0281] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for inspecting airport perimeter fencing, characterized in that, Applications in perimeter inspection robots include: Based on a pre-established segmented map of the airport perimeter and a unified time reference, the fence segment number where the perimeter inspection robot is located and the positioning coordinates of the perimeter inspection robot are determined. Under safe conditions, the boundary inspection robot is controlled to enter a safe approach state and its body posture is adjusted; and the current motion parameters of the boundary inspection robot are collected simultaneously. Based on the current action parameters, determine the effective displacement of the fence along the loading direction; The fence status characteristic quantity is determined based on the effective displacement, and the fence defect category is output based on the deviation between the fence status characteristic quantity and the normal fence response benchmark corresponding to the fence segment number. Image information of the area inside and outside the fence is collected, target recognition is performed on the image information to determine the target category of the fence intrusion target, and the target trajectory of the fence intrusion target is established; Based on the fence centerline and inner / outer directions of the fence segment associated with the fence intrusion target, a fence centerline determination surface is established. Based on the spatial relationship between the target trajectory and the fence centerline determination surface, it is determined whether the fence intrusion target is located outside the fence, in the fence neighborhood, inside the fence, or has made a valid crossing from the outside to the inside, and the personnel and biological intrusion identification results are output. According to the fence segment number and time window, the physical inspection results corresponding to the fence defect category are associated with the personnel and biological intrusion identification results, and the alarm level is determined based on the confirmed intrusion, physical status or security risk event. According to the alarm level, the alarm evidence package and inspection record will be uploaded to the airport inspection management platform.
2. The method according to claim 1, characterized in that, The step of establishing a fence centerline determination surface based on the fence centerline and inner / outer directions of the fence segment associated with the fence intrusion target includes: Obtain the fence centerline, inner and outer directions, and allowed association range of the fence segment associated with the fence intrusion target stored in the airport perimeter segment map, and establish the fence centerline determination surface based on the fence centerline and the inner and outer directions; The method further includes: When the fence intrusion target is located outside the fence and moves parallel to it, an outer trajectory record is formed; when the fence intrusion target continues to approach the fence, an approach trend record is formed. The fence centerline determination surface, inner and outer markings, and target trajectory are used for identifying spatial relationships within the fence.
3. The method according to claim 1, characterized in that, The method further includes: Establish continuous trajectories for airborne targets in the areas inside and outside the fence, and project the trajectories of the airborne targets onto the fence centerline determination surface, fence segments, and the controlled area inside the fence boundary; When the airborne target enters the vicinity of the fence from the outside of the fence and continues to move towards the inside of the fence, or enters the controlled area and maintains a traceable trajectory, an airborne object intrusion identification result is formed; If the confidence level of the category of the airborne target is insufficient, output "unknown airborne target". If the airborne target is only suspended on the fence mesh, signboard or connecting component and does not form a continuous cross-boundary trajectory from the outside to the inside, the airborne target will be classified as a foreign object suspension inspection.
4. The method according to claim 1, characterized in that, The method further includes: The trajectories of multiple individuals located in the vicinity outside the fence are spatiotemporally aggregated to determine the number of individuals, spatial density, dwell time, consistency of movement direction, changes in distance from the center line of the fence, and approach trends of the multiple individuals. If the multiple individuals do not approach the fence, do not exhibit climbing or crossing behavior, and are not effectively associated with abnormal vibrations in the same fence section, a crowd attention event is output. If any of the multiple persons continuously approaches the fence, exhibits climbing behavior, exhibits crossing behavior, or forms a valid association with the abnormal vibration, an intrusion identification event is re-established centered on the trajectory of that person.
5. The method according to claim 1, characterized in that, The method further includes: When the outline of the fence intrusion target continuously overlaps with the fence area, the target trajectory continuously changes along the fence height direction and conforms to climbing characteristics, and the fence vibration sensor of the same fence segment or adjacent fence segments allowed by calibration outputs abnormal output within the associated time window, the climbing probability and identification quality mark of the fence intrusion target are formed, and the trajectory identifier, fence segment number, route coordinates, shortest distance, fence overlap continuity, height direction change characteristics, movement direction characteristics, vibration correlation and input quality of the fence intrusion target are recorded. If the vibration quality of the fence is substandard, the fence intrusion target is severely obstructed, or the inner and outer markings are unstable or the trajectory is not continuous enough, the personnel climbing will be output for verification. If the target of the fence intrusion is a person and is located near the fence, the fence inspection robot is controlled to maintain a safe distance for observation, without approaching, loading, or contacting the fence; after the person leaves and the safety conditions are restored, a physical status check of the fence is arranged.
6. The method according to claim 1, characterized in that, The method further includes: The target trajectory before, during, and after crossing is checked for continuity to rule out apparent crossing caused by temporary obstruction, viewpoint switching, positioning jump, or unstable fence segment affiliation. When the target trajectory continuously enters the inner side of the fence from the outside of the fence, and the fence intrusion target remains inside the fence and effectively crosses the center line determination surface of the fence, output the personnel intrusion identification result or the biological intrusion identification result. If the intrusion target only enters the vicinity of the fence for a short time, does not effectively cross the center line of the fence, cannot be continuously tracked, or the quality of the evidence is insufficient, the result to be verified is output.
7. The method according to claim 1, characterized in that, The method further includes: Based on the visible appearance information, trajectory, location and current time of the fence intrusion target, the consistency is verified with the valid work permit information issued for the current task, and the verification result of the authorized operator is generated. If the fence intrusion target is located within the permitted work area and the valid time period, the visible appearance information is consistent with the parameter package, and the valid work permit information matches the current task, output that the authorized operator has passed verification; If the valid work permit information is valid but the appearance features, trajectory continuity, visibility ratio, observation distance, or nighttime imaging quality are insufficient, output "Authorized operator pending review"; If the fence intrusion target is located in a non-permitted area, the operation time is inconsistent, the permit status is invalid, or there is insufficient visible evidence, the personnel identity is output as pending verification, and the personnel intrusion identification and event classification rules continue to apply.
8. The method according to claim 1, characterized in that, The method further includes: The current day / night mode is determined based on environmental perception quality, unified time reference, and current parameter package. In daytime mode, visible light cameras are the primary identification source, while infrared or thermal imaging devices and lidar are secondary identification sources. In nighttime mode, passive vibrations from infrared or thermal imaging devices, lidar, and fence vibration sensors are the primary identification source, while visible light low-light images are secondary identification sources. For each recognition event, record the day / night mode, available modal set, image quality, thermal imaging effectiveness, point cloud registration quality, positioning quality, time synchronization quality, vibration quality, trajectory continuity, fence segment binding quality, model version, and reason for degradation. In night mode, the event level that has met the same evidence conditions will not be reduced due to modality switching; when the complete recognition conditions cannot be met only because some modalities are unavailable, the recognition results and quality marks of the currently available modalities will be output; when the evidence is insufficient, the results to be verified will be output.
9. The method according to claim 1, characterized in that, The process of determining the fence segment number where the fence inspection robot is located and its positioning coordinates based on a pre-established airport perimeter segment map and a unified time reference includes: Multiple positioning results are obtained, and the reliability of each positioning result is determined; wherein, the reliability of the positioning result is configured to be updated based on the residuals across positioning sources; The positioning results that pass the consistency test are dynamically weighted and fused to output the current parameters of the perimeter inspection robot.
10. The method according to claim 9, characterized in that, The method further includes: If the credibility drops to a first threshold, the perimeter inspection robot is controlled to decelerate and increase the safe distance between itself and the fence. If the credibility drops to a second threshold, the perimeter inspection robot is controlled to stop and reposition itself in place; wherein the second threshold is less than the first threshold. In the event of a repositioning failure, the perimeter inspection robot is controlled to return along the confirmed trajectory or wait for remote takeover.
11. The method according to claim 1, characterized in that, The step of performing target recognition on the image information, determining the target category of the fence intrusion target, and establishing the target trajectory of the fence intrusion target includes: Image information of the area inside and outside the fence is acquired using a visible light camera, infrared or thermal imaging equipment; Using a detection and tracking model trained on airport scene samples, identify at least one target among people, dogs, birds, and other animals in the image information; A continuous trajectory is established for each identified target, and target information is output; wherein, the target information includes: target category, detection and tracking confidence, position, speed, direction of movement, closest distance to the center line of the fence, inner and outer markings, associated fence segment number, route coordinates, geographic coordinates, unified timestamp, and identification quality mark.
12. The method according to claim 1, characterized in that, The method further includes: Based on the relationship between the target trajectory and the center line of the fence, determine the climbing probability and crossing probability of the target trajectory; The detection results of the fence intrusion target include at least one of the following: target category, climbing probability, crossing probability, and vibration anomaly probability corresponding to the fence vibration signal.
13. The method according to claim 12, characterized in that, The step of determining the climbing probability and crossing probability of the target trajectory based on the relationship between the target trajectory and the center line of the fence includes: A crossing determination surface is constructed using the centerline of the fence and the inner and outer directions; A concern event is generated when the fence intrusion target is located outside the fence and moves parallel to it; If the fence intrusion target continues to approach the fence, increase the attention level of the fence intrusion target; The climbing probability is formed when the outline of the fence intrusion target overlaps with the fence area and the height or position change of the target trajectory conforms to climbing characteristics; The crossing probability is formed when the target trajectory continuously enters the inner side of the fence from the outside of the fence and meets the spatial and duration conditions.
14. The method according to claim 1, characterized in that, The step of controlling the perimeter inspection robot to enter a safe approach state and adjust its body posture when safety conditions are met includes: The real-time positioning coordinates are projected onto the target fence segment to determine the lateral and heading errors of the fence inspection robot relative to the inspection offset line. The fence normal is determined according to the fence direction associated with the fence segment number. Under the condition that the safety conditions are met, the side touch detection module is deployed and the body posture is adjusted so that the side touch detection module contacts the fence along the fence normal. If the aforementioned safety conditions are not met, the perimeter inspection robot is controlled to reposition itself or to collect preset data.
15. The method according to claim 1, characterized in that, in, The current action parameters include: contact force, displacement of the side touch detection module, and base displacement; determining the effective displacement of the fence along the loading direction based on the current action parameters includes: Based on the equivalent flexibility of the side touch detection module and the contact force, the displacement of the side touch detection module and the displacement of the base are compensated to obtain the effective displacement; The base displacement is obtained by jointly estimating the kinematics, inertial measurement attitude, and short-time positioning increment of the boundary inspection robot.
16. The method according to claim 15, characterized in that, in, The step of determining the fence state characteristic quantity based on the effective displacement includes: Based on the contact force and the effective displacement, determine the equivalent stiffness and normalized hysteresis, and determine the residual displacement after unloading; The vibration signal of the fence after removing the base movement was decomposed into time and frequency to determine the energy ratio and vibration attenuation coefficient of each frequency band. The fence state characteristic quantity is determined based on at least one of the equivalent stiffness, the normalized hysteresis, the residual displacement, the energy ratio, the vibration attenuation coefficient, and the spatial consistency of adjacent detection points.
17. The method according to claim 16, characterized in that, The method further includes: Determine the baseline group according to the fence parameters corresponding to the fence segment number; The fence state characteristics are standardized using the median and median absolute deviation of the reference group; If there are insufficient benchmark samples, the sample should be returned to the next higher benchmark group. If the data quality is unsatisfactory, stop outputting the fence defect category.
18. The method according to claim 17, characterized in that, The process involves determining the fence state characteristic quantity based on the effective displacement, and outputting the fence defect category based on the deviation between the fence state characteristic quantity and the normal fence response benchmark corresponding to the fence segment number, including: If at least one of the following conditions is met: the equivalent stiffness is reduced relative to the normal fence response benchmark, the normalized hysteresis is increased, the vibration attenuation coefficient is decreased, or the energy ratio exceeds the normal frequency band, and the anomaly occurs at the mesh detection point, the fence defect category is determined to be mesh loosening. If a sudden change occurs during the loading process, and it is consistent with the sudden change at the same location in the preset state, the fence defect category is determined to be broken wire or detached weld; wherein, the preset state includes: visual linear gap, missing connection point or adjacent repeated action; If the residual displacement exceeds the residual displacement threshold and the zero-force position of repeated touch actions remains in the same direction of deviation, the fence defect category is determined to be mesh deformation. If low-frequency displacements in the same direction occur on both sides of the same column or at multiple adjacent detection points, the compensated base displacement is insufficient to explain the low-frequency displacements, and the low-frequency energy or attenuation characteristics exceed the normal reference of the column, the fence defect category is determined to be column instability.
19. The method according to claim 1, characterized in that, The process involves associating the physical inspection results corresponding to the fence defect categories with the personnel and biological intrusion identification results according to the fence segment number and time window, and determining the alarm level based on the confirmed intrusion, physical status, or security risk events, including: The event association set is determined by the personnel and biological intrusion identification results, the physical inspection results corresponding to the fence defect category, and the fence vibration anomaly results that are the same fence segment number or belong to adjacent fence segments that are allowed by the calibration and whose time difference is within the time window. Based on the event association set, the confirmed events are classified into personnel or biological intrusion events, personnel climbing events, high-intensity impact events, abnormal physical condition of fence events, and access risk events. Based on the confirmed event type, generate an observation, attention, warning, or emergency alarm.
20. The method according to claim 19, characterized in that, The method further includes: If any submodule in the personnel and biological intrusion identification, the fence defect category diagnosis, or the fence vibration detection is unavailable or has substandard identification quality, the identification result of the available mode or the available submodule will be downgraded to participate in the event association, and the reason for the missing, the reason for the downgrade, and the quality mark will be recorded. When the safety conditions do not allow for active touch control, the alarm level is determined solely based on visual, thermal imaging, and passive vibration evidence.
21. The method according to claim 1, characterized in that, The perimeter inspection robot is a multi-legged wheeled robot, which includes a body, multiple mechanical legs, and a drive wheel located at the end of each mechanical leg; the method further includes: A local elevation map is established based on the depth of the laser signal and visual signal, and the terrain score of the terrain where the multi-legged wheeled robot is located is determined by combining the current motion parameters and terrain parameters of the multi-legged wheeled robot. The movement pattern of the multi-legged wheeled robot is determined based on the terrain score. The movement modes include: wheeled high-speed driving mode, low-speed wheeled driving mode, wheel-foot coordinated obstacle crossing mode, and safe stop mode.
22. The method according to claim 1, characterized in that, in, The airport perimeter segmentation map includes multiple fence segments; Each fence segment has a unique number, and each fence segment is associated with fence information and area information. The fence segment number, route coordinates, location coordinates, and unified timestamp are stored together.
23. The method according to claim 1, characterized in that, The method further includes: In the task initialization state, the normal appearance reference, mesh geometric reference, connection point reference, sign position and posture reference, perimeter box external closure reference, pole and box external geometric reference, basic neighborhood reference, normal fence response reference and applicable parameter version are loaded according to the fence segment number. Project the current positioning coordinates onto the target fence section or facility route coordinates, and form an observation pose consistent with the fence normal, inspection offset line and safety distance; if the positioning reliability, lateral distance uncertainty, fuselage attitude, lateral space, support margin or target safety distance do not meet the requirements, maintain the non-contact detection mode of vision, thermal imaging, lidar and passive vibration. The system acquires visible light images, infrared or thermal imaging images, laser point clouds, inertial measurement attitude, fence vibration signals and operating status from multiple observation poses for the same fence section or facility, and registers the laser point cloud with the corresponding geometric reference, outputting the registration result and registration quality mark. If a candidate anomaly is detected by visual, thermal imaging, laser point cloud, or passive vibration, and the safety conditions are met and there is no risk of personnel, animals, or unknown targets near the fence, the side touch detection module is invoked to verify the detection points that can be safely touched. The inspection results of the fence and its ancillary facilities are bound to the task number, event category, fence segment number, route coordinates, geographical coordinates, unified timestamp, sensor source, input quality mark, benchmark version, parameter package version, handling status, and record status.
24. The method according to claim 23, characterized in that, The fence and its ancillary facilities include the fence body, and the inspection items of the fence body include at least one of the following: fence damage inspection, fence corrosion, fence screws falling off, foreign objects hanging on the fence, stability of the signboard and other appearance abnormalities, and fence perimeter box. The fence damage inspection is triggered when the visible light image shows abnormal grid continuity, linear gaps or disconnection at the connection point, the laser point cloud shows fence outline deviation, partial missing or abnormal opening, or the passive vibration shows abnormal impact or continuous vibration. The image features are associated with the point cloud features, and the output is pending verification when there is only a single mode and occlusion, shadow, vegetation, reflection or point cloud missing cannot be ruled out. The corrosion of the fence is triggered when the current visible light texture, color, surface mottled features, or appearance changes of the connection parts are abnormal relative to the normal appearance benchmark of the same fence type and the same environmental range. The touch results are only used as auxiliary evidence of the structural state and should not replace the judgment of the appearance of corrosion. The inspection of fallen fence screws is triggered when the visible appearance of the connection point is inconsistent with the reference of the connector, the relative position of adjacent components changes, or the point cloud has local geometric anomalies. The inspection is also triggered by equivalent stiffness, normalized hysteresis, residual displacement and vibration characteristics to distinguish between fallen connectors, loose connectors and general obstruction. The foreign object suspension fence inspection forms a foreign object suspension event when the visible object still maintains spatial association with the fence mesh or signage neighborhood under multiple observation poses. When the target is in a continuous moving state and the target trajectory crosses the fence centerline determination surface from the outside to the inside, it switches to airborne object intrusion identification. The perimeter inspection of the enclosure only identifies the outer outline of the enclosure, the door cover plane, the closed boundary, the tilt status, the connection relationship of the surrounding enclosure, the basic neighborhood, and visible external damage. It does not open the enclosure, touch the electrical components inside the enclosure, or read the internal electrical status.
25. The method according to claim 23, characterized in that, The fence and its ancillary facilities include monitoring poles, light poles, intermediate power distribution boxes, and power distribution boxes inside the monitoring station. The complete inspection of the monitoring pole is used to identify the missing, tilted, hanging or obviously detached parts of the pole, equipment suspension position, speaker, supplementary light, sound and light alarm, lightning rod and its visible external installation status. The complete inspection of the light pole is used to identify abnormalities in the center line of the light pole, the position of the light body, the posture of the light body, the suspension status of the light body, the tilt of the pole and the relative position of the foundation, and does not interpret the absence of thermal response as a failure or missing light fixture. The inspection of the intermediate distribution box and the distribution box in the monitoring station is only used to identify the outline of the external box, the closed status of the door, the tilt of the box, the water accumulation in the vicinity of the foundation and the relative position of the foundation. The box is not opened, the internal wiring is not touched, the electrical status inside the box is not read, and the external temperature map of the box, the appearance of the door or the condition of the foundation is not used to presume an internal power distribution fault. If the equipment is obstructed, the point cloud registration is insufficient, the nighttime image quality is substandard, or the location uncertainty affects the facility's ownership, the status of the corresponding facility will be output as pending verification.
26. The method according to claim 23, characterized in that, The method further includes: Identify potential candidates for water accumulation in low-lying areas, around box foundations, and around pole foundations; In cases where water surface reflection, absorption, occlusion, or sparse point cloud results in unreliable laser results, the quality defect is recorded, and the missing point cloud is not interpreted as water accumulation. If the multi-view visual or thermal imaging evidence is consistent with the baseline of the low-lying area and the point cloud or ground geometry supports the corresponding positional relationship, a water accumulation candidate is formed; otherwise, the low-lying area status is output for verification. Identify the continuous geometric offsets of the edges, adjacent posts, mesh baselines, and ground elevation relationships of the fence foundation, pole foundation, and box foundation relative to historical benchmarks; When a continuous and unidirectional geometric offset is observed, and the registration quality, positioning quality, and benchmark matching conditions meet the requirements, a candidate for foundation settlement is formed; among them, the effective displacement, equivalent stiffness, normalized hysteresis, vibration energy, vibration attenuation coefficient, residual displacement, and spatial consistency obtained by active touch control serve as auxiliary structural evidence. When local ground accessibility does not meet requirements, maintain a safe distance to collect evidence from a distance.
27. The method according to claim 1, characterized in that, The method further includes: In the task initialization state, load the route coordinate range, allowed passage boundary, test direction, road segment benchmark, passable scoring parameter package and safe stop strategy of the test segment, and write the actual road segment width, slope range, effective length, surface condition, load-bearing diameter, total test mass of robot, environmental conditions, test direction, parameter version and expected passage method into the test record. Local elevation maps are created using lidar, and combined with data from visible light, infrared or thermal imaging, inertial measurement units, wheel speed sensors, and foot contact force sensors, input evidence is obtained for slope, roughness, ditch width and depth, wheel-ground adhesion status, aircraft stability margin, landing feasibility, and slip rate. For ground attributes that cannot be reliably estimated by existing sensor quality, the credibility of the drivability score is reduced. The vehicle maintains wheeled driving as long as the passability score meets the conditions for high-speed wheeled driving and the slip rate, attitude, and boundary constraints meet the requirements; it switches to low-speed wheeled driving when the slope, roughness, or slip rate increases but the continuous rolling condition is still met; it switches to wheel-foot cooperative mode when there are ditches, steps, insufficient continuous rolling, or more reliable landing planning; and it switches to safe stop when the wheel-ground adhesion state, fuselage stability margin, landing feasibility, positioning quality, lateral boundary, or sensor quality do not meet the safety conditions; the switching adopts a hysteresis rule. If the prescribed actions of entering, continuously passing, controlled deceleration or braking, and leaving the road segment are completed within the test direction and the permitted passage boundary, and no stability margin, lateral boundary, continuous slippage, critical sensor failure, or safety control overstepping conditions are triggered, it is recorded as passing; if the road segment conditions exceed the calibration range or the sensor quality is insufficient, a safe stop or no valid conclusion is selected.
28. The method according to claim 1, characterized in that, The method further includes: Establish scenario-based functional verification records according to the scenario, time mode, detection object, expected output and handling status; In cases where a valid normal baseline is missing, the current environment is outside the sample, the sensor quality does not meet the requirements, the parameter package is expired, the model version is mismatched, or the event location cannot be reliably bound to the fence segment, the output will be "no valid conclusion" or "to be verified", and the available original evidence will be saved.
29. A perimeter inspection robot, characterized in that, include: The robot body comprises a robot controller, a positioning module, a touch detection module, an environmental perception module, an intrusion detection module, a fusion decision module, and a communication module. The robot controller is connected to the positioning module, the touch detection module, the environmental perception module, the intrusion detection module, the fusion decision module, and the communication module. The positioning module is used to determine the fence segment number where the fence inspection robot is located and the positioning coordinates of the fence inspection robot based on a pre-established airport perimeter segment map and a unified time reference. The robot controller is used to control the perimeter inspection robot to enter a safe approach state and adjust its body posture when safety conditions are met; and to simultaneously collect the current motion parameters of the perimeter inspection robot. The touch detection module is used to determine the effective displacement of the fence along the loading direction based on the current action parameters; The environmental perception module is used to collect image information and determine the detection result of fence intrusion targets; The environmental perception module is also used to collect image information of the area inside and outside the fence, perform target recognition on the image information, determine the target category of the fence intrusion target, and establish the target trajectory of the fence intrusion target; The intrusion identification module is used to establish a fence centerline determination surface based on the fence centerline and the inner and outer directions of the fence segment associated with the fence intrusion target. Based on the spatial relationship between the target trajectory and the fence centerline determination surface, it determines whether the fence intrusion target is located outside the fence, in the fence neighborhood, inside the fence, or has made a valid crossing from the outside to the inside, and outputs personnel and biological intrusion identification results. The fusion decision module is used to associate the physical inspection results corresponding to the fence defect category with the personnel and biological intrusion identification results according to the fence segment number and time window, and determine the alarm level based on the confirmed intrusion, physical status or security risk event. The communication module is used to upload alarm evidence packages and inspection records to the airport inspection management platform according to the alarm level.
30. The boundary inspection robot according to claim 29, characterized in that, The robot body is a multi-legged wheeled body; The multi-legged wheeled body includes a body, multiple mechanical legs, and a drive wheel located at the end of each mechanical leg.
