A robot autonomous navigation method and system based on airport perimeter

Abstract

The application discloses a kind of robot autonomous navigation method and system based on airport perimeter, belong to robot vision navigation technical field.The method includes: multi-modal perception pre-processing output standardization data;Perimeter perception direct network end-to-end output airport perimeter feature and lane basic feature;Based on BEV technology, generate two-dimensional grid map containing perimeter contour and obstacle;From the map, extract perimeter feature point set, generate virtual lane with safety constraint by parameterized curve fitting and relative perimeter translation and time series smoothing;Based on virtual lane, generate control instruction to realize lane keeping and perimeter driving;Real-time state is collected simultaneously to front end each layer and carries out closed-loop optimization and safety monitoring.The application discards GNSS and static map dependence, solves the problem that robot positioning fails under strong electromagnetic interference of airport perimeter, inspection blind area and poor environmental adaptability, realizes high-precision, high-continuity perimeter autonomous navigation.

Description

Technical Field

[0001] This invention relates to the field of robot vision navigation technology, specifically to a robot autonomous navigation method and system based on airport perimeter. Background Technology

[0002] With the deepening of smart airport construction, the need for routine, high-precision inspection of airport perimeter fencing, as the first line of defense for aviation safety, is becoming increasingly urgent. Currently, most autonomous navigation solutions for airport perimeter inspection robots employ a combination of GPS / RTK positioning and static SLAM maps, or visual obstacle avoidance and target point navigation. However, these solutions face several technical bottlenecks in practical operations. First, the airport perimeter often experiences strong electromagnetic interference from tall fencing, dense vegetation, and radar / communication base stations, leading to GNSS signal loss or a significant increase in multipath errors, causing robot positioning failure and making it highly susceptible to getting lost or deviating from the inspection route. Second, traditional SLAM navigation relies on pre-built high-precision point cloud maps, but the airport perimeter environment changes frequently (temporary construction, vegetation growth, snow cover, perimeter deformation, etc.), causing static maps to quickly become invalid. Furthermore, updating and maintaining high-precision maps for long-distance perimeters is extremely costly. In addition, existing pure vision navigation mostly adopts an unconstrained mode of "obstacle avoidance plus target point", which lacks structured guidance similar to a motor vehicle lane. When the robot patrols a long straight boundary, it is prone to lateral drift, or even creates a blind spot in monitoring due to excessive obstacle avoidance, thus losing the core value of inspection. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide a robot autonomous navigation method and system based on airport perimeter boundaries. This method solves the problems of robot positioning failure, blind spots in inspection, and poor environmental adaptability under strong electromagnetic interference at airport perimeter boundaries, and achieves high-precision and highly continuous autonomous navigation along the boundary.

[0004] To achieve the above objectives, the embodiments of this invention provide the following technical solutions:

[0005] This application provides a robot autonomous navigation method based on airport perimeter, comprising the following steps: S1, real-time acquisition of multimodal perception data in front of the robot based on a multimodal perception preprocessing layer, and preprocessing to output standardized perception input data; S2, outputting perimeter feature data and lane basic feature data end-to-end from the perception input data through a perimeter perception direct network designed based on the airport perimeter scenario; S3, generating a two-dimensional grid map of the airport perimeter using BEV technology based on the perimeter feature data and lane basic feature data, the two-dimensional grid map of the airport perimeter including perimeter outline and obstacle information; S4, extracting a continuous feature point set of the airport perimeter from the two-dimensional grid map of the airport perimeter, generating a virtual lane that maintains a preset safe distance from the airport perimeter through parametric curve fitting and relative airport perimeter translation, and performing temporal smoothing processing on the generated virtual lane; S5, generating motion control commands based on the virtual lane to control the robot to maintain lane and drive close to the perimeter.

[0006] Further, S1 specifically includes: S11, the multimodal perception preprocessing layer performs distortion correction and adaptive illumination enhancement processing on the acquired image; S12, the multimodal perception preprocessing layer performs spatiotemporal synchronization processing on the multi-source perception data; S13, the multimodal perception preprocessing layer performs anomaly filtering and standardization processing on the preprocessed data, and outputs the standardized perception input data.

[0007] Furthermore, the boundary-aware direct network is trained using a multi-task joint loss function, the expression of which is: ;in For obstacle detection regression loss, For instance segmentation loss, To enforce the continuity of the airport perimeter outline, the airport perimeter continuity loss, The regression loss is for lane-based feature points. These are the weighting coefficients.

[0008] Further, S4 specifically includes: S41, transforming the airport boundary feature point set in the image coordinate system to the robot's local coordinate system using the virtual lane generation and smoothing layer to obtain airport boundary feature points in the local coordinate system; S42, based on the airport boundary feature points in the local coordinate system, the virtual lane generation and smoothing layer offsets outward by a preset boundary safety distance along the vertical direction of the robot's forward movement, and generates a base point set for the virtual lane center guide line; S43, the virtual lane generation and smoothing layer offsets by a preset lateral safety distance on both sides of the normal direction of the virtual lane center guide line, and generates a base point set for the left and right safety boundary lines; S44, the virtual lane generation and smoothing layer performs parametric curve fitting on the base point set of the virtual lane center guide line and the base point set of the left and right safety boundary lines to generate a virtual lane; S45, the virtual lane generation and smoothing layer performs temporal smoothing processing on the virtual lanes generated in multiple consecutive frames to obtain a smoothed virtual lane.

[0009] Furthermore, the parameterized curve fitting in S44 uses a cubic Bézier curve, expressed as: ;in, Parameters on the curve The corresponding point coordinates These are the four control points of the Bézier curve. As the starting point of the lane, The end of the lane, and As an intermediate control point, The normalized curve parameter takes values ​​in a closed interval from 0 to 1; and when there are right-angle turns or broken line inflection points in the airport perimeter, a spiral curve is used for the inflection point transition.

[0010] Furthermore, S4 also includes: when an obstacle is detected, the virtual lane generation and smoothing layer performs hierarchical local lane reconstruction based on the size, position and dynamic and static attributes of the obstacle, and generates detour sub-lanes.

[0011] Further, S5 specifically includes: S51, the lane keeping control layer calculates the lateral deviation and heading angle deviation of the robot relative to the virtual lane center guide line; S52, the lane keeping control layer constructs a speed control model based on lane curvature, perception confidence, obstacle distance, and road surface condition to solve for the optimal driving speed, the expression of the speed control model being: ;in, To achieve the optimal driving speed, The preset maximum driving speed, The curvature attenuation factor, , For lane curvature; To perceive the confidence decay factor, The segmentation confidence is the output of the boundary-aware direct network. The obstacle distance attenuation factor. The distance between the robot and the obstacle. The road surface condition attenuation factor is determined based on IMU bump data; S53, the lane keeping control layer generates steering and speed control commands based on the lateral deviation, heading angle deviation and optimal driving speed.

[0012] Furthermore, in step S53, a model predictive control algorithm is used to generate steering and speed control commands, the expressions of which are: ;in, To predict the number of time-domain steps, To control the number of time-domain steps, This is the state deviation vector. , For lateral deviation, For heading angle deviation, For speed deviation, The state weight matrix is... This is a vector of control variable rates of change, including the rate of change of steering angle and the rate of change of acceleration. To control the weight matrix, The rate of change of the distance between the robot and the airport perimeter. Weighting for fluctuations in boundary distance. For relaxation factor weights, These are slack variables.

[0013] Furthermore, after S5, it also includes: S6, through the state feedback and safety monitoring layer, collecting the robot's real-time motion state and perception state, feeding it back to at least one of the layers in S1-S5, so as to perform closed-loop optimization of each layer, while monitoring the safety status and triggering graded early warning and emergency intervention.

[0014] Accordingly, this application also provides a robot autonomous navigation system based on airport perimeter, comprising: a multimodal perception preprocessing layer, used to collect multimodal perception data in front of the robot in real time based on the multimodal perception preprocessing layer, and perform preprocessing to output standardized perception input data; a perimeter perception direct network layer, connected to the multimodal perception preprocessing layer, used to output perimeter feature data and lane basic feature data end-to-end from the perception input data through a perimeter perception direct network designed based on the airport perimeter scenario; a BEV perimeter 2D map generation layer, connected to the perimeter perception direct network layer, used to generate an airport perimeter 2D grid map based on the perimeter feature data and lane basic feature data using BEV technology, the airport perimeter 2D grid map including perimeter outline and obstacle information; and a virtual lane generation and smoothing layer, connected to the BEV perimeter 2D map generation layer, used to extract the airport perimeter from the airport perimeter 2D grid map. A continuous set of feature points is used to generate virtual lanes that maintain a preset safe distance from the airport boundary through parametric curve fitting and translation relative to the airport boundary. The generated virtual lanes are then subjected to temporal smoothing. A lane-keeping control layer, connected to the virtual lane generation and smoothing layer, is used to generate motion control commands based on the virtual lanes to control the robot to maintain its lane and drive close to the boundary. A state feedback and safety monitoring layer, connected to the multimodal perception preprocessing layer, the boundary perception direct network layer, the BEV boundary 2D map generation layer, the virtual lane generation and smoothing layer, and the lane-keeping control layer, is used to collect the robot's real-time motion and perception states and feed them back to at least one of these layers for closed-loop optimization. Simultaneously, it monitors the safety status and triggers tiered warnings and emergency interventions.

[0015] The beneficial effects of this invention are as follows: By directly outputting boundary features and lane basic features through an end-to-end neural network specifically designed for airport perimeter scenarios, and combining it with bird's-eye view modeling technology to generate a structured virtual lane strongly bound to the boundary, the robot is provided with clear navigation constraints and guidance. It can maintain continuous and stable navigation capabilities in environments with no or weak GNSS, such as strong electromagnetic interference and signal obstruction. At the same time, it effectively solves the problems of lateral drift and blind spots formed by traditional pure vision navigation, which are far from the boundary, and significantly improves the coverage and safety of airport perimeter inspection. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a robot autonomous navigation method based on an airport perimeter provided in this application embodiment;

[0017] Figure 2This is a schematic diagram of a robot autonomous navigation system based on an airport perimeter, provided as an embodiment of this application. Detailed Implementation

[0018] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0019] In this invention, the terms "system" and "network" are used interchangeably. "Multiple" refers to two or more; therefore, in this invention, "multiple" can also be understood as "at least two." "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / ", unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, it should be understood that in the description of this invention, terms such as "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or order.

[0020] Example 1:

[0021] like Figure 1 As shown in the figure, this application provides a robot autonomous navigation method based on airport perimeter, including the following steps: S1, real-time acquisition of multimodal perception data in front of the robot based on a multimodal perception preprocessing layer, and preprocessing to output standardized perception input data; S2, outputting perimeter feature data and lane basic feature data end-to-end from the perception input data through a perimeter perception direct network designed based on the airport perimeter scenario; S3, generating a two-dimensional grid map of the airport perimeter using BEV technology based on the perimeter feature data and lane basic feature data, the two-dimensional grid map of the airport perimeter including perimeter outline and obstacle information; S4, extracting a continuous feature point set of the airport perimeter from the two-dimensional grid map of the airport perimeter, generating a virtual lane that maintains a preset safe distance from the airport perimeter through parametric curve fitting and relative airport perimeter translation, and performing temporal smoothing processing on the generated virtual lane; S5, generating motion control commands based on the virtual lane to control the robot to maintain lane and drive close to the perimeter.

[0022] In another possible embodiment, multimodal perception data in front of the robot is first acquired in real time through a multimodal perception preprocessing layer. This multimodal perception data specifically includes RGB images acquired by a forward-looking wide-angle camera, angular velocity and acceleration data acquired by an IMU (Inertial Measurement Unit), and 3D point cloud data acquired by an optional LiDAR. These raw data undergo preprocessing operations such as distortion correction, illumination enhancement, spatiotemporal synchronization, and anomaly filtering to remove noise, distortion, and invalid information, unify the spatiotemporal reference and format of the data, and output standardized perception input data. The standardized perception input data is then input into a boundary perception direct network specifically designed for airport boundary scenarios. This network adopts an end-to-end architecture that integrates a lightweight Transformer and CNN, eliminating the need for a separate instance segmentation step in traditional solutions. It can directly extract boundary feature data, including boundary masks and contour key points, as well as basic lane feature data for subsequent virtual lane generation, from the input data. It can also output the position, size, and dynamic / static attribute information of obstacles. Based on the boundary feature data and lane basic feature data output from the network, BEV (Browser-Electric Vehicle) bird's-eye view technology is used to convert the two-dimensional features on the image plane into three-dimensional spatial features from a bird's-eye view, generating a two-dimensional grid map of the airport boundary containing clear boundary outlines, obstacle positions, and attribute information. A continuous set of boundary feature points is extracted from the generated airport boundary two-dimensional grid map. These feature points are smoothed using parametric curve fitting, and the fitted curve is shifted outwards by a preset safe distance relative to the airport boundary to generate a virtual lane that maintains a fixed safe distance from the boundary. Simultaneously, temporal smoothing is performed on the virtual lanes generated in multiple consecutive frames to eliminate random jitter caused by single-frame detection, resulting in a stable spatiotemporal virtual lane corridor. Finally, based on the generated virtual lane parameters and the robot's real-time motion state, the lateral deviation and heading angle deviation of the robot relative to the virtual lane center guide line are calculated, generating corresponding steering and speed control commands to control the robot to drive smoothly along the virtual lane, achieving precise lane keeping and boundary inspection. Throughout the inspection process, the status feedback and safety monitoring layer collects the robot's motion and perception status in real time, feeds it back to the front-end layers for closed-loop optimization, and monitors the safety status to trigger graded early warnings and emergency interventions.

[0023] By directly outputting boundary features and lane basic features through an end-to-end neural network specifically designed for airport perimeter scenarios, and combining it with bird's-eye view modeling technology to generate structured virtual lanes strongly bound to the boundary, the robot is provided with clear navigation constraints and guidance. It can maintain continuous and stable navigation capabilities in environments with no or weak GNSS, such as strong electromagnetic interference and signal obstruction. At the same time, it effectively solves the problems of lateral drift and blind spots formed by moving away from the boundary in traditional pure vision navigation, and significantly improves the coverage and safety of airport perimeter inspection.

[0024] In this embodiment of the application, S1 specifically includes: S11, the multimodal perception preprocessing layer performs distortion correction and adaptive illumination enhancement processing on the acquired image; S12, the multimodal perception preprocessing layer performs spatiotemporal synchronization processing on the multi-source perception data; S13, the multimodal perception preprocessing layer performs anomaly filtering and standardization processing on the preprocessed data, and outputs the standardized perception input data.

[0025] In another possible embodiment, the multimodal perception preprocessing layer first performs distortion correction on the raw image captured by the camera. Using pre-calibrated camera intrinsic parameters, including focal length, principal point coordinates, and distortion coefficients, it corrects the barrel and pincushion distortions caused by the camera lens itself, obtaining an image with accurate geometric shape. Next, the distortion-corrected image undergoes adaptive illumination enhancement processing, specifically employing the Retinex algorithm combined with histogram equalization. The Retinex algorithm can separate the illumination and reflection components of the image, eliminating the influence of uneven illumination, while histogram equalization can stretch the grayscale range of the image and improve image contrast. This effectively improves the visual effect of images in harsh environments such as low light, backlight, and heavy fog, making boundary features such as barbed wire, posts, and low walls more clearly identifiable. Then, the acquired multi-source sensing data, including images, IMU data, and LiDAR point clouds, undergoes spatiotemporal synchronization processing. Hardware trigger signals synchronize the acquisition times of each sensor, ensuring all sensors acquire data simultaneously. A unified high-precision timestamp is added to each data set to achieve temporal synchronization. Simultaneously, pre-calibrated extrinsic parameters between sensors convert data from different sensors to the robot's body coordinate system, achieving spatial synchronization. Finally, the synchronized multi-source data undergoes anomaly filtering and standardization. Physically valid data that deviates from physical laws is removed, such as sudden negative wheel speeds, steering angles exceeding mechanical limits, or abnormal peaks in IMU data. Data lacking critical information is interpolated or removed. The remaining valid data is converted to a unified format and numerical range, and standardized sensing input data is output to the boundary sensing direct network layer.

[0026] Targeted image preprocessing algorithms effectively improved image clarity and contrast in harsh environments, making boundary features more prominent. Spatiotemporal synchronization of multi-source data ensured precise alignment of data from different sensors in time and space. Abnormal data filtering eliminated abrupt and invalid data that did not conform to physical laws, providing high-quality input data for the subsequent boundary perception direct network, effectively improving the accuracy of boundary feature extraction and the overall stability of the system.

[0027] In this embodiment, the boundary-aware direct network is trained using a multi-task joint loss function, the expression of which is: ;in For obstacle detection regression loss, For instance segmentation loss, To enforce the continuity of the airport perimeter outline, the airport perimeter continuity loss, The regression loss is for lane-based feature points. These are the weighting coefficients.

[0028] In another possible embodiment, the boundary perception direct network employs a multi-task joint loss function for parameter optimization during the training phase. This loss function is composed of a weighted average of four parts: obstacle detection regression loss, instance segmentation loss, airport boundary continuity loss, and lane basic feature point regression loss, expressed as follows: ,in For obstacle detection regression loss, For instance segmentation loss, To enforce the continuity of the airport perimeter outline, the airport perimeter continuity loss, The regression loss is for lane-based feature points. The weights are denoted as . Further, the obstacle detection regression loss uses CIoU loss, which simultaneously considers the overlap area, center distance, and aspect ratio of the predicted and ground truth boxes, effectively improving the detection and localization accuracy of small obstacles such as debris and small equipment along the airport perimeter. The instance segmentation loss uses Dice loss, which addresses the imbalanced sample problem and optimizes the segmentation effect of the perimeter and obstacle masks, enabling the network to clearly distinguish between the perimeter, obstacles, and background areas. The perimeter continuity loss is a loss term specifically designed for the linear topological features of the airport perimeter. By calculating the distance deviation between adjacent key points on the perimeter contour, it penalizes the breakage of the perimeter contour, forcing the network to output continuous perimeter lines. The lane basic feature point regression loss uses mean squared error loss to optimize the localization accuracy of lane pre-feature points, ensuring the accuracy of subsequent virtual lane generation. By adjusting the weights of each loss term, the network focuses more on the perimeter continuity and lane basic feature extraction tasks closely related to boundary-fitting navigation during training, thus obtaining a perception model more suitable for airport perimeter scenarios. In actual reasoning, the trained boundary-aware direct network can simultaneously output the results of the above four tasks without the need for separate branch processing, thus reducing intermediate data processing steps and error accumulation.

[0029] By employing a multi-task joint loss function, the four tasks of obstacle detection, boundary instance segmentation, boundary contour continuity, and lane basic feature extraction are jointly optimized. Through the weight coefficients optimized for airport boundary scenarios, the importance of boundary continuity and lane basic feature extraction is highlighted, effectively improving the continuity of boundary features and the positioning accuracy of lane basic features, reducing boundary gap detection and false detection, and laying a solid foundation for the accurate generation of virtual lanes in the future.

[0030] In this embodiment, S4 specifically includes: S41, the virtual lane generation and smoothing layer transforms the airport boundary feature point set in the image coordinate system to the robot's local coordinate system to obtain the airport boundary feature points in the local coordinate system; S42, the virtual lane generation and smoothing layer, based on the airport boundary feature points in the local coordinate system, offsets outward by a preset boundary safety distance along the vertical direction of the robot's forward movement, and generates a base point set for the virtual lane center guide line; S43, the virtual lane generation and smoothing layer offsets by a preset lateral safety distance on both sides of the normal direction of the virtual lane center guide line, and generates a base point set for the left and right safety boundary lines; S44, the virtual lane generation and smoothing layer performs parametric curve fitting on the base point set of the virtual lane center guide line and the base point set of the left and right safety boundary lines to generate a virtual lane; S45, the virtual lane generation and smoothing layer performs temporal smoothing processing on the virtual lanes generated in multiple consecutive frames to obtain a smoothed virtual lane.

[0031] In another possible embodiment, the virtual lane generation and smoothing layer first receives the boundary feature point set output by the BEV boundary 2D map generation layer. It transforms the boundary feature points from the image coordinate system to the camera coordinate system using pre-calibrated camera extrinsic parameters. Then, combining this with real-time robot pitch and roll angle attitude data acquired by the IMU, it dynamically corrects projection errors caused by ground bumps using a homography matrix. Because the robot encounters uneven road surfaces and potholes during its movement, causing changes in camera attitude, direct coordinate transformation would produce significant projection errors. Dynamic correction effectively eliminates this influence, ultimately accurately transforming the feature points to the robot's local coordinate system, obtaining the boundary feature points in the local coordinate system. Then, based on these boundary feature points in the local coordinate system, a preset safety distance is offset outward along the vertical direction of the robot's movement to generate the base point set for the virtual lane center guide line. A preset lateral safety distance is then offset to both sides of the center guide line's normal direction to generate the base point sets for the left and right safety boundary lines. Finally, parametric curve fitting is performed on the base point sets of the center guide line and the left and right safety boundary lines to generate smooth and continuous virtual lane lines. Finally, the virtual lane parameters generated in multiple consecutive frames are subjected to a three-level temporal smoothing process. The first level uses inter-frame Kalman filtering, which establishes a uniform motion state equation with the lane centerline control point as the state variable to eliminate random noise from single-frame detection. The second level uses sliding window temporal fusion, which sets a fixed-length sliding window and performs weighted least squares fitting on the lane parameters of multiple frames within the window to ensure the continuity of lanes between frames. The third level uses global topological constraints of the airport perimeter, which corrects the cumulative directional deviation generated during long-distance travel based on the pre-stored overall orientation and inflection point information of the perimeter, and finally obtains stable and reliable virtual lane parameters.

[0032] Coordinate transformation and dynamic error compensation improve the positioning accuracy of boundary feature points in the robot's local coordinate system. Direct translation based on boundary features to generate virtual lanes ensures the accuracy and stability of the boundary distance. Parametric curve fitting makes the lane smoother and more continuous. Three-level temporal smoothing effectively eliminates random noise and jitter in single-frame detection, greatly improving the stability of the virtual lane and enabling the robot to drive smoothly and accurately along the lane.

[0033] In this embodiment of the application, the parameterized curve fitting in S44 adopts a cubic Bézier curve, the expression of which is: ;in, Parameters on the curve The corresponding point coordinates These are the four control points of the Bézier curve. As the starting point of the lane, The end of the lane, and As an intermediate control point, The normalized curve parameter takes values ​​in a closed interval from 0 to 1; and when there are right-angle turns or broken line inflection points in the airport perimeter, a spiral curve is used for the inflection point transition.

[0034] In another possible embodiment, when performing parametric curve fitting on the virtual lane base point set, the topological structure type of the boundary is first determined based on the distribution of boundary feature points. For conventional straight line segments and gentle curve segments, a cubic Bézier curve is used for fitting, expressed as follows: The four control points of the curve are adaptively solved using the sliding window least squares method, where... Parameters on the curve The corresponding point coordinates These are the four control points of the Bézier curve. As the starting point of the lane, The end of the lane, and As an intermediate control point, The normalized curve parameter takes values ​​in a closed interval from 0 to 1, where This represents the starting point of the lane corresponding to the current robot position. The lane terminus within the forward look-ahead distance. and As an intermediate control point, The normalized curve parameters are set within a closed interval from 0 to 1. During the solution process, the continuity of the second derivative of the curve and the minimum rate of curvature change are ensured, thereby generating a smooth virtual lane and making the robot's steering more stable. When a right-angle turn or a polygonal inflection point is detected in the boundary, a cubic Bézier curve is no longer used; instead, a spiral curve is employed for the inflection point transition. The curvature of the spiral curve is proportional to its length, enabling a smooth transition from a straight line to a curve and back to a straight line. This ensures continuous change in lane curvature, preventing abrupt steering changes and sudden changes in boundary distance when the robot passes through inflection points. This guarantees that the robot can smoothly and accurately pass through various complex boundary structures while reducing mechanical wear on the steering mechanism.

[0035] By employing differentiated curve fitting methods for different boundary topologies, cubic Bézier curves are used for conventional straight / curved boundaries to ensure the smoothness and continuity of the lane. Spiral curves are used for transitions at right-angle turns or inflection points of broken lines, so that the lane curvature changes linearly with the length, avoiding the steering impact caused by sudden curvature changes, and improving the robot's driving smoothness and boundary-fitting accuracy under complex boundary topologies.

[0036] In this embodiment of the application, S4 further includes: when an obstacle is detected, the virtual lane generation and smoothing layer performs hierarchical local lane reconstruction based on the size, position and dynamic and static attributes of the obstacle, and generates a detour sub-lane.

[0037] In another possible embodiment, when the virtual lane generation and smoothing layer detects an obstacle from the BEV boundary 2D map, it first classifies the obstacle's risk based on its size, degree of intrusion relative to the virtual lane, and dynamic / static attributes. For low-risk obstacles that are completely outside the virtual lane and do not affect the robot's movement, the original virtual lane remains unchanged, and the robot continues to drive normally along the original route. For medium-risk obstacles that partially intrude into the virtual lane but still have sufficient detour space, a local detour sub-lane is generated based on the parameters of the original virtual lane and the obstacle's position and size information. The curvature change rate of the detour sub-lane is strictly limited within a reasonable range to ensure smooth robot movement. At the same time, the distance deviation between the robot and the boundary is strictly controlled during the detour. After the detour, the robot automatically and smoothly returns to the original virtual lane and continues boundary inspection. For high-risk obstacles that completely block the virtual lane and cannot be safely detoured, a parking warning is immediately triggered, controlling the robot to slow down and stop. Simultaneously, detailed information about the obstacle and its current location is uploaded to the remote monitoring platform to remind operators to remotely view and handle the situation to avoid collisions. For dynamic obstacles such as birds and inspection personnel, their movement trajectories are predicted, and detour lanes are dynamically updated to ensure detour safety.

[0038] By conducting graded risk assessments based on the size, location, and dynamic / static attributes of obstacles, different handling strategies are adopted for obstacles of different risk levels. Low-risk obstacles do not affect normal driving, medium-risk obstacles generate local detour sub-lanes to ensure that the vehicle can quickly return to the original lane and maintain boundary accuracy after detour, and high-risk obstacles trigger parking warnings. This approach ensures driving safety while minimizing the impact on boundary inspection and avoiding the creation of blind spots in the inspection process.

[0039] In this embodiment, S5 specifically includes: S51, the lane keeping control layer calculates the lateral deviation and heading angle deviation of the robot relative to the virtual lane center guide line; S52, the lane keeping control layer constructs a speed control model based on lane curvature, perception confidence, obstacle distance, and road surface condition to solve for the optimal driving speed, wherein the expression of the speed control model is: ;in, To achieve the optimal driving speed, The preset maximum driving speed, The curvature attenuation factor, , For lane curvature; To perceive the confidence decay factor, The segmentation confidence is the output of the boundary-aware direct network. The obstacle distance attenuation factor. The distance between the robot and the obstacle. The road surface condition attenuation factor is determined based on IMU bump data; S53, the lane keeping control layer generates steering and speed control commands based on the lateral deviation, heading angle deviation and optimal driving speed.

[0040] In another possible embodiment, the lane keeping control layer first receives virtual lane parameters output by the virtual lane generation and smoothing layer, and real-time robot status data output by the status feedback and safety monitoring layer. It then calculates the lateral deviation of the robot's center of mass in the direction perpendicular to the lane center guide line, and the deviation of the robot's current heading angle from the tangent direction of the lane center guide line at its current position. Next, based on the curvature of the virtual lane, the segmentation confidence score output by the boundary perception direct network, the distance between the robot and obstacles, and the road surface condition determined based on IMU bump data, a multi-dimensional speed control model is constructed, comprehensively considering the influence of various factors on the driving speed. The expression is: .in, To achieve the optimal driving speed, The preset maximum driving speed, The curvature attenuation factor, , For lane curvature; To perceive the confidence decay factor, The segmentation confidence is the output of the boundary-aware direct network. The obstacle distance attenuation factor. The distance between the robot and the obstacle. The road condition attenuation factor, determined based on IMU bump data, is used to determine the robot's speed. The curvature attenuation factor decreases as lane curvature increases, automatically reducing the robot's speed on high-curvature curves. The perception confidence attenuation factor decreases as perception confidence decreases; when perception confidence falls below a certain threshold, the robot significantly reduces its speed to ensure safety. The obstacle distance attenuation factor decreases as the distance between the robot and obstacles decreases, allowing for low-speed movement and sufficient reaction time at close range. The road condition attenuation factor is calculated based on vertical acceleration data collected by the IMU; the greater the road bumps, the smaller the attenuation factor and the lower the speed. The optimal speed under the current conditions is calculated by multiplying the maximum speed by each attenuation factor. Finally, based on the calculated lateral deviation, heading angle deviation, and optimal speed, corresponding steering angle and speed control commands are generated and sent to the robot's actuators to control the robot to precisely travel along the virtual lane.

[0041] A multi-dimensional linkage speed control model was constructed, which can dynamically adjust the driving speed in real time according to various environmental and state factors, thereby improving driving safety in different scenarios. By accurately calculating the lateral deviation and heading angle deviation of the robot relative to the virtual lane, precise control commands are generated, which greatly improves the accuracy of lane keeping and boundary driving.

[0042] In this embodiment of the application, step S53 uses a model predictive control algorithm to generate steering and speed control commands, the expressions of which are: ;in, To predict the number of time-domain steps, To control the number of time-domain steps, This is the state deviation vector. , This is the lateral deviation. For heading angle deviation, For speed deviation, The state weight matrix is... This is a vector of control variable rates of change, including the rate of change of steering angle and the rate of change of acceleration. To control the weight matrix, The rate of change of the distance between the robot and the airport perimeter. Weighting for fluctuations in boundary distance. For relaxation factor weights, These are slack variables.

[0043] In another possible embodiment, when generating steering and speed control commands, a model predictive control algorithm is employed. First, a state-space equation is constructed based on the single-vehicle kinematic model of the inspection robot. This equation defines the state vector as including the robot's position coordinates, heading angle, speed, lateral deviation, and heading angle deviation. The control input vector includes the steering angle and acceleration. The model parameters are calibrated according to the actual driving characteristics of the airport perimeter inspection robot. Then, an optimization objective function is constructed, expressed as follows: ,in, To predict the number of time-domain steps, To control the number of time-domain steps, This is the state deviation vector. , This is the lateral deviation. For heading angle deviation, For speed deviation, The state weight matrix is... This is a vector of control variable rates of change, including the rate of change of steering angle and the rate of change of acceleration. To control the weight matrix, The rate of change of the distance between the robot and the airport perimeter. Weighting for fluctuations in boundary distance. For relaxation factor weights, The variables are relaxation variables, where the state deviation penalty term minimizes the pose deviation between the robot and the virtual lane, with a focus on strengthening the penalty for lateral deviation; the control quantity smoothing penalty term limits the rate of change of steering angle and acceleration to ensure smooth control and reduce mechanical wear; the boundary distance fluctuation penalty term forces the robot to maintain a stable safe distance from the boundary to avoid it straying too far from the boundary; and the soft constraint relaxation term ensures the feasibility of the optimization problem. During the optimization process, hard constraints on the upper and lower limits of the boundary distance and the obstacle safety distance are added to ensure that the robot's driving state is always within a safe range. By rolling the solution to the optimal control sequence in the finite time domain, executing only the first control variable in the sequence at each step, and then resolving the control sequence in the next time domain based on the new state data, real-time closed-loop optimization control is achieved, resulting in higher control accuracy and smoothness than pure tracking algorithms.

[0044] By employing model predictive control algorithms, it is possible to simultaneously optimize state deviation, control quantity smoothness, and boundary distance fluctuation within a finite time domain. Furthermore, hard constraints on boundary distance and obstacle safety distance are added, which not only significantly improves the control accuracy of lane keeping and boundary driving but also ensures the smoothness of control quantities, making the robot drive more smoothly and guaranteeing driving safety from the control layer.

[0045] In this embodiment of the application, after S5, it also includes: S6, collecting the robot's real-time motion state and perception state through the state feedback and safety monitoring layer, feeding it back to at least one of the layers in S1-S5, so as to perform closed-loop optimization of each layer, while monitoring the safety status and triggering graded early warning and emergency intervention.

[0046] In another possible embodiment, throughout the robot's inspection task, the state feedback and safety monitoring layer collects the robot's motion and perception state data in real time via the IMU (Inertial Measurement Unit), wheel speed encoder, vision sensor, and actuator feedback interface. The motion state data includes the robot's position, heading angle, speed, acceleration, steering angle, and wheel speed. The perception state data includes the confidence level of boundary feature extraction, obstacle detection results, and sensor operating status. The collected multi-source data is synchronized with high precision, with synchronization errors controlled within milliseconds. Abnormal data is then removed through physical rationality verification. An extended Kalman filter algorithm is used to fuse and filter the multi-source state data, eliminating sensor noise and sampling jitter, and outputting smooth and stable core robot state parameters. This real-time state data is fed back to at least one of the following layers: the multimodal perception preprocessing layer, the boundary perception direct network layer, the BEV (Boundary Elevated Vehicle) boundary 2D map generation layer, the virtual lane generation and smoothing layer, and the lane keeping control layer. This allows the layer to dynamically adjust the virtual lane generation parameters and control commands according to the robot's actual operating state, achieving closed-loop optimization throughout the entire process. Simultaneously, the collected status data is analyzed and monitored in real time. The inspection status is divided into three levels: normal execution, early warning, and abnormal termination. When the status indicator approaches the safety threshold, an early warning is triggered, and the system operating parameters are automatically adjusted, such as reducing the driving speed and increasing the filtering intensity, to avoid risks. The early warning information is also sent to the monitoring platform. When the status indicator exceeds the safety threshold, emergency intervention is triggered. The robot is immediately controlled to brake and the inspection operation is terminated. At the same time, detailed abnormal information and fault codes are uploaded to the monitoring platform to notify the operation and maintenance personnel to handle the situation in a timely manner and ensure the safe operation of the system.

[0047] The system achieves closed-loop optimization of the entire process through status feedback, enabling the navigation system to adapt to the robot's actual status changes in real time, continuously adjust the virtual lane generation parameters and control commands, and improve the accuracy and stability of navigation. At the same time, through a hierarchical safety monitoring mechanism, it can promptly detect various abnormal situations and trigger corresponding early warnings and intervention measures, effectively ensuring the safe operation of the system.

[0048] Example 2:

[0049] Reference Figure 2This application also provides a robot autonomous navigation system based on airport perimeter, comprising: a multimodal perception preprocessing layer, used to collect multimodal perception data in front of the robot in real time based on the multimodal perception preprocessing layer, and preprocess the data to output standardized perception input data; a perimeter perception direct network layer, connected to the multimodal perception preprocessing layer, used to output perimeter feature data and lane basic feature data end-to-end from the perception input data through a perimeter perception direct network designed based on the airport perimeter scenario; a BEV perimeter 2D map generation layer, connected to the perimeter perception direct network layer, used to generate an airport perimeter 2D grid map based on the perimeter feature data and lane basic feature data using BEV technology, the airport perimeter 2D grid map including perimeter outline and obstacle information; and a virtual lane generation and smoothing layer, connected to the BEV perimeter 2D map generation layer, used to extract airport perimeter connections from the airport perimeter 2D grid map. Continuing with the feature point set, a virtual lane is generated by parametric curve fitting and relative airport boundary translation, maintaining a preset safe distance from the airport boundary. The generated virtual lane is then subjected to temporal smoothing. A lane-keeping control layer, connected to the virtual lane generation and smoothing layer, is used to generate motion control commands based on the virtual lane to control the robot to maintain its lane and drive close to the boundary. A state feedback and safety monitoring layer, connected to the multimodal perception preprocessing layer, the boundary perception direct network layer, the BEV boundary 2D map generation layer, the virtual lane generation and smoothing layer, and the lane-keeping control layer, is used to collect the robot's real-time motion and perception states and feed them back to at least one of these layers for closed-loop optimization, while simultaneously monitoring safety status and triggering tiered warnings and emergency interventions.

[0050] In another possible embodiment, the system adopts a closed-loop layered architecture, consisting of, from bottom to top, a multimodal perception preprocessing layer, a boundary perception direct network layer, a BEV boundary 2D map generation layer, a virtual lane generation and smoothing layer, a lane keeping control layer, and a state feedback and safety monitoring layer that communicates with all the above layers. The multimodal perception preprocessing layer, as the system's perception input, consists of a forward-looking wide-angle camera, optional LiDAR, an IMU inertial measurement unit, and a matching preprocessing algorithm module. It collects various perception data in front of the robot in real time and preprocesses them, pushing standardized perception input data to the boundary perception direct network layer at a fixed frame rate. The boundary perception direct network layer is the core perception module of the system, an end-to-end neural network specifically designed for airport boundary scenarios. After receiving standardized perception data, it directly outputs boundary feature data, obstacle information, and lane basic feature data to the BEV boundary 2D map generation layer. The BEV boundary 2D map generation layer, combined with pre-calibrated camera intrinsic and extrinsic parameters, converts the input feature data into a bird's-eye view 2D grid map. After annotating the boundary outline and obstacle information, it outputs to the virtual lane generation and smoothing layer. The virtual lane generation and smoothing layer extracts boundary feature points from the 2D grid map, generating a virtual lane with safety constraints. After time-series smoothing, it outputs to the lane keeping control layer. The lane keeping control layer generates steering and speed control commands based on virtual lane parameters and the robot's real-time status, sending them to the robot's actuators to control the robot to perform boundary-following inspections. The status feedback and safety monitoring layer collects the robot's operating and perception status in real time, pushing status data to all front-end layers, triggering closed-loop optimization at each layer, and performing full-process safety monitoring. Based on abnormal situations, it triggers tiered warnings and emergency interventions to ensure safe system operation. All layers interact with each other through standardized API interfaces, ensuring clear and efficient data flow and forming a complete closed-loop control chain.

[0051] By constructing a hierarchical closed-loop system architecture, each layer has a clear function and works closely together. From multimodal perception, boundary feature extraction, BEV map generation, virtual lane generation to lane keeping control and safety monitoring, a closed-loop control link is formed throughout the entire process. This completely eliminates the dependence on GNSS signals and static maps, and enables high-precision and high-continuity autonomous navigation in environments with no or weak GNSS, greatly improving the system's stability, reliability and environmental adaptability.

[0052] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention.

[0053] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not describe the various possible combinations separately.

[0054] Furthermore, various different implementations of the present invention can be combined arbitrarily, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed in the present invention.

Claims

1. A robot autonomous navigation method based on airport perimeter, characterized in that, Includes the following steps: S1. Based on the multimodal perception preprocessing layer, collect multimodal perception data in front of the robot in real time, perform preprocessing, and output standardized perception input data; S2. Through the boundary perception direct network designed based on the airport boundary scenario, boundary feature data and lane basic feature data are output end-to-end from the perception input data. S3. Based on the boundary feature data and lane basic feature data, generate a two-dimensional grid map of the airport boundary using BEV technology. The two-dimensional grid map of the airport boundary includes the boundary outline and obstacle information. S4. Extract the continuous feature point set of the airport boundary from the two-dimensional grid map of the airport boundary, generate a virtual lane that maintains a preset safe distance from the airport boundary through parametric curve fitting and relative airport boundary translation, and perform time-series smoothing processing on the generated virtual lane. S5. Generate motion control commands based on the virtual lane to control the robot to maintain lane position and drive close to the boundary.

2. The robot autonomous navigation method based on airport perimeter as described in claim 1, characterized in that, S1 specifically includes: S11. The acquired image is subjected to distortion correction and adaptive illumination enhancement processing by the multimodal sensing preprocessing layer. S12. The multimodal sensing preprocessing layer performs spatiotemporal synchronization processing on the multi-source sensing data; S13. The multimodal perception preprocessing layer performs anomaly filtering and standardization on the preprocessed data and outputs the standardized perception input data.

3. The robot autonomous navigation method based on airport perimeter as described in claim 1, characterized in that, The boundary-aware direct network is trained using a multi-task joint loss function, the expression of which is: ； in For obstacle detection regression loss, For instance segmentation loss, To enforce the continuity of the airport perimeter outline, the airport perimeter continuity loss, The regression loss is for lane-based feature points. These are the weighting coefficients.

4. The robot autonomous navigation method based on airport perimeter as described in claim 1, characterized in that, S4 specifically includes: S41. The set of airport boundary feature points in the image coordinate system is transformed to the robot local coordinate system by the virtual lane generation and smoothing layer to obtain the airport boundary feature points in the local coordinate system. S42. Based on the airport boundary feature points in the local coordinate system, the virtual lane generation and smoothing layer offsets outward by a preset boundary safety distance in the direction perpendicular to the robot's forward direction, and generates a set of base points for the virtual lane center guide line. S43. The virtual lane is generated by offsetting a preset lateral safety distance from the smoothing layer on both sides of the normal direction of the virtual lane center guide line, and a set of base points for the left and right safety boundary lines is generated. S44. The virtual lane generation and smoothing layer performs parametric curve fitting on the base point set of the virtual lane center guide line and the base point set of the left and right safety boundary lines to generate a virtual lane. S45. The virtual lanes generated in multiple consecutive frames are subjected to temporal smoothing by the virtual lane generation and smoothing layer to obtain smoothed virtual lanes.

5. The robot autonomous navigation method based on airport perimeter as described in claim 4, characterized in that, The parameterized curve fitting in S44 uses a cubic Bézier curve, expressed as follows: ； in, Parameters on the curve The corresponding point coordinates These are the four control points of the Bézier curve. As the starting point of the lane, The end of the lane, and As an intermediate control point, The normalized curve parameter takes values ​​in a closed interval from 0 to 1. Furthermore, when there are right-angle turns or broken line inflection points in the airport perimeter, spiral curves are used for the transition at the inflection points.

6. The robot autonomous navigation method based on airport perimeter as described in claim 1, characterized in that, The S4 further includes: when an obstacle is detected, the virtual lane generation and smoothing layer performs hierarchical local lane reconstruction based on the size, position and dynamic and static attributes of the obstacle, and generates detour sub-lanes.

7. The robot autonomous navigation method based on airport perimeter as described in claim 1, characterized in that, S5 specifically includes: S51. The lane keeping control layer calculates the lateral deviation and heading angle deviation of the robot relative to the virtual lane center guide line; S52. The lane keeping control layer constructs a speed control model based on lane curvature, perception confidence, obstacle distance, and road surface condition to solve for the optimal driving speed. The expression of the speed control model is: ； in, To achieve the optimal driving speed, The preset maximum driving speed, The curvature attenuation factor, , For lane curvature; To perceive the confidence decay factor, The segmentation confidence is the output of the boundary-aware direct network. The obstacle distance attenuation factor. The distance between the robot and the obstacle. The road surface condition attenuation factor is determined based on IMU bump data; S53. The lane keeping control layer generates steering and speed control commands based on the lateral deviation, heading angle deviation, and optimal driving speed.

8. The robot autonomous navigation method based on airport perimeter as described in claim 7, characterized in that, In S53, a model predictive control algorithm is used to generate steering and speed control commands, the expression of which is: ； in, To predict the number of time-domain steps, To control the number of time-domain steps, This is the state deviation vector. , This is the lateral deviation. For heading angle deviation, For speed deviation, The state weight matrix is... This is a vector of control variable rates of change, including the rate of change of steering angle and the rate of change of acceleration. To control the weight matrix, The rate of change of the distance between the robot and the airport perimeter. Weighting for fluctuations in boundary distance. For relaxation factor weights, These are slack variables.

9. The robot autonomous navigation method based on airport perimeter as described in claim 1, characterized in that, Following S5 are: S6. Through the status feedback and safety monitoring layer, the robot's real-time motion status and perception status are collected and fed back to at least one of the layers in S1-S5 to perform closed-loop optimization of each layer, while monitoring the safety status and triggering graded early warnings and emergency interventions.

10. A robot autonomous navigation system based on airport perimeter, characterized in that, The robot autonomous navigation method based on airport perimeter as described in any one of claims 1 to 9 includes: The multimodal perception preprocessing layer is used to collect multimodal perception data in front of the robot in real time, perform preprocessing, and output standardized perception input data. The boundary perception direct network layer, connected to the multimodal perception preprocessing layer, is used to output boundary feature data and lane basic feature data end-to-end from the perception input data through the boundary perception direct network designed based on the airport boundary scenario. The BEV boundary 2D map generation layer is connected to the boundary perception direct network layer. It is used to generate a 2D grid map of the airport boundary using BEV technology based on the boundary feature data and lane basic feature data. The 2D grid map of the airport boundary includes the boundary outline and obstacle information. The virtual lane generation and smoothing layer is connected to the BEV boundary 2D map generation layer. It is used to extract the continuous feature point set of the airport boundary from the airport boundary 2D grid map, generate a virtual lane that maintains a preset safe distance from the airport boundary through parametric curve fitting and relative airport boundary translation, and perform time-series smoothing processing on the generated virtual lane. The lane keeping control layer, connected to the virtual lane generation and smoothing layer, is used to generate motion control commands based on the virtual lane to control the robot to maintain its lane and drive along the boundary. The status feedback and safety monitoring layer is connected to the multimodal perception preprocessing layer, the boundary perception direct network layer, the BEV boundary 2D map generation layer, the virtual lane generation and smoothing layer, and the lane keeping control layer. It is used to collect the robot's real-time motion and perception status and feed it back to at least one of the multimodal perception preprocessing layer, the boundary perception direct network layer, the BEV boundary 2D map generation layer, the virtual lane generation and smoothing layer, and the lane keeping control layer for closed-loop optimization of each layer, while monitoring the safety status and triggering graded early warnings and emergency interventions.

Images