Unmanned aerial vehicle adaptive flight control method in GPS-denied environment under bridge lower part

Abstract

The application discloses a bridge lower part GPS rejection environment unmanned aerial vehicle adaptive flight control method, belongs to high-precision positioning technical field, and its technical scheme main points include: based on multi-source sensor navigation data establishes bridge lower part environment physical effect coupling database and bridge lower part characteristic map; according to environment prediction result obtains preloaded physical effect compensation parameter set and initializes double SLAM fusion positioning system; through covariance reconstruction algorithm and multi-physical constraint factor graph optimization method obtains unmanned aerial vehicle's optimization position estimation, physical credibility grid and real-time effect compensation degree; through double-path closed loop calibration updates physical effect coupling database and obtains enhanced navigation positioning result, and the application realizes unmanned aerial vehicle high-precision, high-robust positioning in bridge lower part self-excitation interference environment through physical effect modeling compensation and branch decision fusion mechanism, solves the traditional fusion decision failure problem caused by sensor performance covariant.

Description

Technical Field

[0001] This invention relates to the field of high-precision positioning technology, and more specifically to an adaptive flight control method for unmanned aerial vehicles (UAVs) in a GPS-denied environment under a bridge. Background Technology

[0002] With the rapid development of intelligent inspection of infrastructure such as bridges, the demand for high-precision navigation and positioning for autonomous drone inspections is increasing. In the substructure of bridges, the enclosed space and severe obstruction create a typical GPS-denied environment, rendering conventional navigation and positioning methods that rely on satellite signals completely ineffective. Stable and reliable high-precision positioning has become a core prerequisite for drones to carry out inspection operations safely and efficiently.

[0003] Existing UAV positioning solutions in GPS-denied environments mostly employ lidar or visual simultaneous localization and mapping (SLAM) technologies as core methods. Some solutions also attempt to improve positioning robustness through simple multi-sensor fusion. However, in the complex environment beneath bridges, the coupled effects of multiple physical effects, such as laser multipath reflection caused by dense structures, extreme lighting variations due to light leakage from bridge deck joints, and dust generated by UAV rotor airflow disturbances, can simultaneously degrade the observation data quality of both lidar and visual sensors, leading to feature matching instability and co-variation in sensor performance.

[0004] Traditional multi-sensor fusion positioning methods are generally designed based on the assumption that sensor performance is independent. This makes it difficult to cope with the synchronous degradation of sensors caused by the coupling of multiple physical effects. It is also difficult to continuously output pose estimation results that meet the accuracy requirements of bridge inspection. These methods generally suffer from technical defects such as insufficient positioning accuracy, poor environmental adaptability, and large drift during long-term operation. They cannot guarantee high-precision and stable positioning of UAVs in GPS-denied environments under bridges. Therefore, existing technologies have shortcomings. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention aims to provide an adaptive flight control method for unmanned aerial vehicles (UAVs) in GPS-denied environments under bridges. This method establishes a local feature map under the bridge that integrates multiple physical effect parameters and pre-initializes a dual-SLAM system. It employs branch decision-making driven by the collaborative compensation of active optical probes and vision, as well as laser scanning matching effects. The method performs fusion calculations based on covariance reconstruction of decoupled physical effects and optimization of multi-physical constraint factor maps. Real-time enhancement of the database and positioning results is achieved through dual-path closed-loop calibration. This enables high-precision and robust positioning of UAVs under self-excited interference environments such as extreme lighting, multipath reflection, and self-contamination under bridges. It resolves the fundamental contradiction of sensor performance covariance leading to the failure of traditional fusion decision-making mechanisms, and improves the positioning reliability and environmental adaptability of UAVs during autonomous inspections in GPS-denied environments.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention provides an adaptive flight control method for unmanned aerial vehicles (UAVs) in a GPS-denied environment under bridge substructure, comprising:

[0008] We acquire navigation data from multiple sources of UAV sensors, and obtain a physical effect coupling database and a local feature map under the bridge by modeling optical transmission characteristics, analyzing polarization characteristics and resolving multipath reflection paths.

[0009] By acquiring real-time GPS navigation signal strength and combining it with local feature maps under the bridge, an initial dual-SLAM fusion positioning system is obtained through environmental prediction using a bridge structure model and a physical effect propagation model.

[0010] Based on the initialization dual SLAM fusion localization system, the optimal pose estimate of the UAV is obtained through physical effect compensation, covariance reconstruction algorithm and multi-physical constraint factor graph optimization method.

[0011] Based on optimized pose estimation, the physical effect coupling database is updated through dual-path closed-loop calibration to obtain enhanced navigation and positioning results.

[0012] As a further improvement of the present invention, the acquisition of multi-source sensor navigation data of the UAV, and the generation of a physical effect coupling database and a local feature map under the bridge through optical transmission characteristic modeling, polarization characteristic analysis and multipath reflection path analysis; includes:

[0013] Based on navigation data from multiple sources, optical transmission function parameters are obtained by modeling optical transmission characteristics, and multipath reflection characteristic parameters are obtained by analyzing multipath reflection paths.

[0014] Polarization characteristic parameters are obtained based on multi-source sensor navigation data through Stokes vector calculation and polarization characteristic analysis.

[0015] Based on optical transfer function parameters, polarization characteristic parameters, and multipath reflection characteristic parameters, a physical effect coupled database and a local feature map under the bridge are constructed.

[0016] As a further improvement of the present invention, the method of obtaining real-time GPS navigation signal strength, combining it with a local feature map under the bridge, and performing environmental prediction through a bridge structure model and a physical effect propagation model to obtain an initialized dual-SLAM fusion positioning system includes:

[0017] Based on real-time GPS navigation signal strength, the distance between the UAV and the blind zone boundary is obtained by comparing signal strength thresholds and using a distance calculation model;

[0018] Based on the pre-set bridge structure model and real-time environmental data, environmental prediction results are obtained through the under-bridge ray tracing model and the physical effect propagation model.

[0019] Based on the environmental prediction results, a preloaded set of physical effect compensation parameters is obtained by matching and retrieving from the physical effect coupling database.

[0020] Based on the preloaded physical effect compensation parameter set and the local feature map under the bridge, an initialized dual SLAM fusion positioning system is obtained through parallel startup and parameter loading processes.

[0021] As a further improvement of the present invention, the initialization-based dual SLAM fusion localization system obtains the optimized pose estimate of the UAV through physical effect compensation, covariance reconstruction algorithm and multi-physical constraint factor graph optimization method, including:

[0022] Based on quantum lidar and vision sensor, the optically compensated synchronously acquired image is obtained by performing physical effect compensation and inverse filtering compensation through optical transfer function cooperative inversion algorithm.

[0023] Based on synchronously acquired images after optical compensation, pollution level estimation and material classification results are obtained through a decoupling model of polarization characteristics and pollution.

[0024] The matching confidence of the lidar point cloud data and the local feature map under the bridge is evaluated in real time, and a branch decision is performed based on the matching confidence to obtain a temporary localization result;

[0025] Based on the temporary positioning results, the optimized pose estimation, physical credibility grid, and real-time effect compensation degree are obtained by using the covariance reconstruction algorithm and the multi-physics constraint factor graph optimization method.

[0026] As a further improvement of the present invention, the step of performing branch decision based on matching confidence to obtain temporary localization results includes:

[0027] When the matching confidence level is greater than or equal to a preset matching confidence threshold, the laser-synchronized localization and mapping (LSRT) system is used as the primary localization source, with visual LSRT as the auxiliary verification system, to obtain a first temporary localization result, and / or...

[0028] When the matching confidence is less than the preset matching confidence threshold, the weight of the visual synchronous localization and mapping system is increased, and the polarization and material decoupling model is activated to enhance visual features, thus obtaining a second temporary localization result.

[0029] As a further improvement of the present invention, the method of obtaining the optically compensated synchronously acquired image based on quantum lidar and a vision sensor through a collaborative inversion algorithm of optical transfer function for physical effect compensation and inverse filtering compensation includes:

[0030] The quantum lidar is used as an active optical probe to emit coded light pulses, and the visual sensor is used in conjunction to acquire synchronously collected images.

[0031] Based on the coded light pulses and synchronously acquired images, the light transmission equation in the under-bridge environment is solved and inverted using an optical transfer function collaborative inversion algorithm to obtain the real-time optical transfer function;

[0032] Based on the real-time optical transfer function, the synchronously acquired image is subjected to inverse filtering to obtain an optically compensated synchronously acquired image.

[0033] As a further improvement of the present invention, the synchronously acquired image based on optical compensation obtains pollution level estimation and material classification results through a polarization characteristic and pollution decoupling model, including:

[0034] Polarization information is obtained by calculating the Stokes vector of the feature region based on the synchronously acquired image after optical compensation.

[0035] Based on multi-source sensor navigation data, the polarization information is decomposed into inherent polarization characteristics and additional polarization effects through a polarization characteristic and pollution decoupling model, and the decomposition result is obtained.

[0036] Based on the decomposition results, the pollution level is estimated by quantifying the additional polarization effect, and the material classification result is obtained by analyzing and judging the inherent polarization characteristics.

[0037] As a further improvement of the present invention, the step of obtaining optimized pose estimation, physical reliability grid, and real-time effect compensation degree based on temporary positioning results through covariance reconstruction algorithm and multi-physics constraint factor graph optimization method includes:

[0038] Based on the pollution level estimation and material classification results, the decoupled sensor observation noise covariance matrix is ​​calculated using a covariance reconstruction algorithm that decouples physical effects.

[0039] Using the temporary positioning result as the initial value, and combining the decoupled sensor observation noise covariance matrix, the optimized pose estimate is obtained through the multi-physics constraint factor graph optimization method.

[0040] Based on the covariance of the optimized pose estimation, a physical credibility grid is obtained through three-dimensional mesh mapping and uncertainty propagation method;

[0041] Based on the pollution level estimation and optical transfer function collaborative inversion algorithm, the real-time effect compensation degree is obtained through weighted fusion.

[0042] As a further improvement of the present invention, the step of updating the physical effect coupling database based on optimized pose estimation and obtaining enhanced navigation and positioning results through dual-path closed-loop calibration includes:

[0043] Based on optimized pose estimation, when the UAV flies close to the pre-stored key structural features, the positioning drift is corrected by structural feature calibration through dual-path closed-loop calibration and the first calibration result is obtained.

[0044] Based on optimized pose estimation, when airflow disturbances are detected that cause flight attitude fluctuations, the positioning deviation is corrected by dynamic disturbance calibration through dual-path closed-loop calibration and a second calibration result is obtained.

[0045] Based on the first and second calibration results, the optical transfer function model and the correlation model between polarization characteristics and material properties are refitted and cross-validated online to obtain the cross-validation results.

[0046] The physical effect coupling database is updated based on the cross-validation results, and the optimized pose estimation is recalculated based on the updated physical effect coupling database to obtain the enhanced navigation and positioning results.

[0047] As a further improvement of the present invention, the multi-source sensor navigation data includes at least lidar point cloud data acquired by quantum lidar, visual image data acquired by visual sensors, and inertial measurement unit data acquired by inertial measurement unit.

[0048] This invention utilizes a physical effect coupled database and a local feature map under a bridge, constructed from navigation data from multiple sensors. It initializes a dual-SLAM fusion positioning system through environmental prediction and pre-compensation parameter loading. The fusion positioning solution is performed based on a branch decision driven by laser scanning matching effects and a covariance reconstruction algorithm decoupled from physical effects. Furthermore, it combines dual-path closed-loop calibration using structural features and dynamic disturbances to continuously enhance the database and positioning results. This achieves high-precision and robust positioning of UAVs under extreme lighting conditions, multipath reflections, and self-contamination interference under bridges. It resolves the fundamental contradiction of sensor performance covariance leading to the failure of traditional multi-sensor fusion weight allocation and decision-making mechanisms, improving the positioning reliability, environmental adaptability, and mission safety of UAVs in complex GPS-denied environments. Attached Figure Description

[0049] Figure 1 This is a flowchart of the method steps of the present invention;

[0050] Figure 2 A flowchart outlining the steps involved in constructing a physical effect coupled database and a local feature map under the bridge.

[0051] Figure 3 A flowchart outlining the steps to obtain the initialized dual SLAM fusion positioning system;

[0052] Figure 4 A flowchart illustrating the steps involved in GPS rejection blind spots, physical effect compensation, and fusion positioning.

[0053] Figure 5 A flowchart illustrating the steps involved in dual-path closed-loop calibration and enhanced navigation and positioning. Detailed Implementation

[0054] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0055] Identical parts are indicated by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, while the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific part, respectively.

[0056] The term "and / or" in the following text is merely a description of 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 " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0057] like Figure 1 As shown, this invention provides an adaptive flight control method for unmanned aerial vehicles (UAVs) in a GPS-denied environment under bridge substructure, comprising:

[0058] To acquire multi-source sensor navigation data of UAVs, the multi-source sensor navigation data includes at least lidar point cloud data acquired by quantum lidar, visual image data acquired by visual sensors, and inertial measurement unit data acquired by inertial measurement unit. Based on the multi-source sensor navigation data, through optical transmission characteristic modeling, polarization characteristic analysis and multipath reflection path analysis, a physical effect coupling database and a local feature map of the bridge sub-environment are established.

[0059] The system acquires real-time GPS navigation signal strength and calculates the distance between the UAV and the blind zone boundary. Before entering the blind zone, it performs environmental prediction based on the bridge structure model and the physical effect propagation model. Based on the environmental prediction results, it obtains a pre-loaded set of physical effect compensation parameters and initializes a dual SLAM fusion positioning system with quantum lidar and visual sensors as the core.

[0060] Within the GPS-denied blind zone, a quantum lidar is used as an active optical probe to work in conjunction with a visual sensor to compensate for physical effects. Branch decisions are executed based on the laser scanning matching effect. Through a covariance reconstruction algorithm with physical effect decoupling and a multi-physical constraint factor graph optimization method, a fusion positioning solution is performed to obtain the optimized pose estimation, physical reliability grid, and real-time effect compensation degree of the UAV.

[0061] Based on optimized pose estimation, the physical effect coupling database is updated through dual-path closed-loop calibration based on structural features and dynamic disturbances to obtain enhanced navigation and positioning results.

[0062] The multi-source sensor navigation data of the UAV consists of raw observation data synchronously collected by its onboard sensor system after the UAV takes off from the GPS effective area on the bridge. This includes at least lidar point cloud data characterizing the spatial geometry of the bridge substructure acquired by quantum lidar, visual image data containing texture and polarization information acquired by a binocular vision camera with a global shutter, and inertial measurement unit (IMU) data reflecting the angular velocity and acceleration of the aircraft. The quantum lidar is a solid-state lidar with high-resolution time-of-flight histogram output capability. Optical transmission characteristic modeling is based on synchronous laser illumination and visual observation, using a method to invert the light field distribution under the bridge by solving the light transmission equation. This method is used to establish a dynamic light field model describing uneven illumination, scattering, and color distortion effects. The study includes: 1) Learning transfer function parameters; 2) Polarization characteristic analysis, which extracts Stokes vectors from visual images to calculate polarization angles and degrees of polarization, used to quantify the polarization optical properties of surface materials; 3) Multipath reflection path analysis, which decomposes the original echo time histogram of lidar to identify and reconstruct multiple reflection paths, used to establish multipath reflection characteristic parameters caused by complex geometries under bridges; 4) A physical effect coupling database for the bridge sub-environment, a digital model library comprehensively storing optical transfer function parameters, polarization characteristic parameters, and multipath reflection characteristic parameters, used to provide data support for real-time physical effect compensation; and 5) A local feature map under the bridge, a digital map constructed based on GPS initial positioning and multi-sensor data fusion, containing geometric information and its corresponding physical effect parameters, used to provide enhanced features for positioning matching in blind areas.

[0063] The real-time GPS navigation signal strength is the GPS carrier-to-noise ratio received by the UAV, acquired in real time by the onboard GPS receiver module; the distance between the UAV and the blind zone boundary is the Euclidean distance calculated based on the UAV's current position and a pre-stored bridge CAD model, obtained through ray casting calculation combining real-time positioning results and the geometric model; the bridge structure model is a pre-digitized 3D computer-aided design model of the bridge, used to provide precise geometric constraints on the space under the bridge, obtained by importing design drawings or 3D scanning data; the physical effect propagation model is a simulation model integrating geometric optics and electromagnetic wave propagation theories, used to predict the propagation of physical effects under specific atmospheric conditions. The propagation path, attenuation, and polarization state changes of light rays under the bridge under certain conditions are analyzed. The environmental prediction results provide a quantitative description of the future illumination distribution, main multipath reflection areas, and polarization change trends within the blind zone. The pre-loaded physical effect compensation parameter set consists of parameter combinations matched and retrieved from the physical effect coupling database based on the environmental prediction results. These parameters include optical transfer function compensation parameters for image inverse filtering, polarization characteristic compensation parameters for feature enhancement, and multipath reflection suppression parameters for suppressing false echoes. The dual SLAM fusion positioning system is a system that operates laser synchronous positioning and mapping and visual synchronous positioning and mapping in parallel and fuses their information tightly.

[0064] Physical effect compensation is a process that uses a pre-loaded set of physical effect compensation parameters to correct the raw laser and visual observation data in real time to offset the effects of environmental degradation, thereby improving the sensor's perception quality in extreme environments. Laser scanning matching effect is a confidence score for feature matching between real-time quantum lidar point cloud data and a local feature map under the bridge, calculated using a point cloud registration algorithm. Branch decision is an intelligent decision-making logic that dynamically selects whether to prioritize laser SLAM or visual SLAM based on a matching confidence threshold, automatically switching to a more reliable positioning source when laser features degrade. The covariance reconstruction algorithm for physical effect decoupling decomposes sensor observation noise into inherent noise, environmental coupling noise, and cross-interference components, and reconstructs its covariance matrix using physical effect estimates, maintaining the effectiveness of the fusion algorithm when multiple sensors simultaneously degrade. The multi-physical constraint factor graph optimization method introduces physical factors such as optical transmission consistency, polarization and material correlation, and multipath reflection geometric constraints into the traditional factor graph optimization framework, ensuring that the pose estimation results are consistent with... The system conforms to physical laws; optimized pose estimation is the optimal position and attitude estimate of the UAV in the local coordinate system obtained after optimization by the above multi-physical constraint factor map; the physical reliability grid is a gridded map formed by mapping the uncertainty of pose estimation to the three-dimensional space under the bridge, with each voxel storing the positioning reliability index of the area; the real-time effect compensation degree is a scalar value calculated by combining the activation intensity of each physical effect compensation module, used to quantify the overall interference caused by the current environment to the perception system; dual-path closed-loop calibration is a parallel calibration mechanism used to correct cumulative drift and instantaneous disturbances respectively, including structural feature calibration to correct long-term positioning drift by comparing with pre-stored accurate structural features, and dynamic disturbance calibration to correct instantaneous positioning deviation by monitoring airflow disturbances with IMU and visual optical flow; the enhanced navigation and positioning result is the UAV pose information with improved accuracy obtained after dual-path closed-loop calibration and model update; the flight control command is the control signal sequence used to directly drive the UAV actuators after solving the rolling time domain optimization problem by the physical fusion predictive control model.

[0065] This embodiment is based on an active perception system composed of high-precision quantum lidar and visual sensors. By constructing a multi-physics effect coupling model, it performs mechanism modeling and predictive compensation for environmental degradation. Based on the branch decision-making mechanism driven by the laser scanning matching effect and the dual-path closed-loop calibration architecture, it achieves a fundamental improvement in the positioning accuracy and robustness of UAVs in extreme environments where GPS is completely denied under bridges. It solves the problem of traditional multi-sensor fusion failure caused by complex physical effects such as drastic changes in illumination, surface contamination, and multipath reflection. This improves the environmental adaptability, flight safety, and operational reliability of UAVs in autonomous inspection missions under bridges.

[0066] Furthermore, this embodiment provides a step-by-step approach to establish a physical effect coupled database and a local feature map of the bridge substructure environment based on multi-source sensor navigation data, through optical transmission characteristic modeling, polarization characteristic analysis, and multipath reflection path resolution. The steps include:

[0067] Based on lidar point cloud data and visual image data from multi-source sensor navigation data, optical transmission function parameters are obtained by optical transmission characteristic modeling, and multipath reflection characteristic parameters are obtained by multipath reflection path analysis.

[0068] Based on visual polarization image data from multi-source sensor navigation data, polarization characteristic parameters are obtained through Stokes vector calculation and polarization characteristic analysis.

[0069] Based on GPS positioning results, lidar point cloud data, and visual image data from multi-source sensor navigation data, and combined with optical transfer function parameters, polarization characteristic parameters, and multipath reflection characteristic parameters, a physical effect coupled database and a local feature map under the bridge containing geometric structures and their corresponding physical effect parameters are constructed.

[0070] The lidar point cloud data consists of a high-resolution 3D spatial point set acquired by a solid-state lidar, which supports nanosecond-level time-of-flight histogram raw data output. This data is obtained by emitting laser pulses and receiving their return signals, providing information on both time and intensity. The visual image data comprises images containing texture and polarization information of the bridge substructure, acquired synchronously by a binocular vision camera. The optical transfer function parameters are dynamic model parameters describing the non-uniformity of illumination distribution, scattering effects, and color distortion during light propagation in the under-bridge environment. The multipath reflection characteristic parameters characterize the path distribution and intensity attenuation of the laser signal due to multiple reflections caused by the complex structure under the bridge. The parameters for the polarization characteristics are as follows: visual polarization image data consists of images containing polarization state information acquired by a binocular vision camera in conjunction with a polarizer array, obtained by acquiring multiple frames of images under different polarization directions and calculating them; Stokes vectors are four-dimensional vectors used to fully describe the polarization state of light waves, used to extract polarization angle and degree of polarization; polarization characteristic parameters are the polarization angle and degree of polarization calculated based on Stokes vectors, and the surface polarization reflection characteristic parameters obtained by fitting a material reflection model; GPS positioning results are the absolute position coordinates obtained by the UAV through the onboard GPS receiver module when it takes off from the effective GPS area on the bridge, used to provide a global reference for the initial map construction.

[0071] Specifically, such as Figure 2As shown, firstly, based on lidar point cloud data and synchronously acquired visual image data, an optical transfer function model is established using an environment-adaptive ray transfer equation inversion algorithm. The laser pulse is considered an active light source, and its time-coded information is synchronized with the camera exposure sequence. By solving the simplified radiative transfer equation, the attenuation coefficient and scattering phase function parameters of the current scene are obtained, thus constructing a point spread function matrix describing the entire process of light from laser emission to camera reception. Simultaneously, the time-flight histogram of the original lidar echo signal is decomposed into multiple peaks. Prior knowledge of the bridge structure is used to constrain the reflection path search, identifying and separating primary and secondary reflection components. The delay time, intensity ratio, and geometric correlation information of each path are extracted to obtain a multipath reflection characteristic parameter set. Then, based on visual polarization image data, the Stokes vector of each preset feature region in the visual polarization image is calculated. The preset feature regions include the lower edge and corners of the bridge. By analyzing the preset feature regions... , , and The intensity observations under four polarization directions are linearly combined to obtain the polarization information. Further, a polarization decomposition network based on surface contamination perception is used to decompose the calculated polarization information into the inherent polarization characteristics of the material and the additional polarization effect caused by surface contamination. The polarization decomposition network employs a lightweight convolutional encoder and decoder structure, using visual polarization image data and corresponding local geometric features of the LiDAR as input. It is trained through supervised learning using a historical bridge-under-contamination dataset. The loss function combines polarization reconstruction error and material classification cross-entropy. The optimizer is Adam, with an initial learning rate of 0.001. The network output includes a contamination degree coefficient and a material category probability distribution.

[0072] Subsequently, the optical transfer function parameters, multipath reflection characteristic parameters, and polarization characteristic parameters are associated with the LiDAR point cloud and visual feature points spatially aligned with GPS positioning results. A voxel-based multi-layer map representation method is used to construct a local feature map under the bridge. Each voxel not only stores geometric attributes such as average position, color, and normal vector, but also stores the corresponding optical transfer function parameter vector, multipath reflection parameter vector, and polarization parameter vector in the form of an association table. Finally, all grid data associated with physical effect parameters are integrated to form the local feature map under the bridge. At the same time, the optical transfer function parameters, polarization characteristic parameters, and multipath reflection characteristic parameters are summarized and stored to construct an indexable, hierarchical physical effect coupled database. This database supports the rapid retrieval of the corresponding physical effect parameter set based on geographical location or feature descriptor. The local feature map under the bridge, as an enhanced digital map containing geometric and physical attributes, provides multi-dimensional feature constraints for positioning and matching of UAVs after entering GPS denied areas.

[0073] For example, this embodiment assumes the substructure of a reinforced concrete box girder bridge with a span of [missing information]. The drone initially lies directly below the midpoint of the bridge surface. Assume the quantum lidar's scanning frequency is... Angular resolution is The frame rate of the binocular vision camera First, a frame containing... is acquired using quantum lidar. Point cloud data of each point Simultaneously acquire a visual image. When modeling optical transmission characteristics, select the image from For each feature point region, a simplified ray transmission equation is constructed using the geometric information of its corresponding spatial points in the point cloud and the laser pulse emission record, incorporating distance attenuation and surface normal constraints. ,in Atmospheric attenuation coefficient, For scattering coefficient, These are the parameters of the surface bidirectional reflection distribution function. For noise, This is an optical transfer function model used to describe the overall attenuation and scattering process of light from laser emission to the camera sensor. The theoretical incident light intensity is calculated based on laser emission power, the inverse square law of distance, and an initial atmospheric attenuation model. The optical transfer function parameter vector is then solved using a gradient descent-based inversion algorithm. ,in The dimension of the parameter vector; the original time histogram of one channel of the quantum lidar in multipath reflection path analysis. ,in This represents the timeline from the start of laser emission. The maximum detectable flight time set for the system typically corresponds to the maximum range of the lidar. An expectation-maximization algorithm is used to fit a time range that includes... A mixture model with Gaussian components, each corresponding to primary reflection, secondary reflection, and noise, is used to extract the mean of the corresponding flight time. Weight and variance This constitutes a multipath reflection characteristic parameter vector. Then, polarization analysis is performed to calculate the Stokes vector for a predefined bolt feature region in the image. ,in Indicates total light intensity. Represents the linear polarization component. Representing the circular polarization component, using the Stokes vector The polarization angle was calculated. and polarization degree ,in The scaling factor is the angle scaling factor, due to the Stokes parameter. , and Proportional, therefore Used to recover the true polarization angle from the arctangent value.

[0074] Then, the obtained polarization angle and degree of polarization, along with the lidar reflection intensity in that region, are analyzed. Input a pre-trained polarization decomposition network, and the network outputs a contamination level estimate. Material classification probability distribution ,in For example, the number of material categories. Corresponding to the four categories of concrete, metal, paint, and contamination layer, the polarization characteristic parameter vector is obtained. Next, map construction is performed, using the point cloud data. In raster resolution Perform voxelization for each non-empty voxel In addition to calculating its center coordinates and average RGB color, it will also calculate the corresponding coordinates of all feature points within its spatial range. Perform average or principal component analysis to obtain the first... Physical effect parameter set of individual elements And create an index, where They represent the first The statistical aggregation result of optical transfer function parameters, multipath reflection characteristic parameters, and polarization characteristic parameters corresponding to all feature points within an individual element. Further exemplarily, when the total light intensity... The observed value is 150. It is 40. When the value is 30, the calculated degree of polarization is 0.33 and the polarization angle is 18.43°. If the polarization decomposition network yields an estimated degree of contamination of [value missing], then [value missing]. , ,because Greater than the preset pollution level threshold This indicates that the area is severely polluted, and the probability distribution is based on material classification. The main material is metal. The parameters used in this embodiment, such as voxel resolution, number of polarization directions, number of network output categories, and number of Gaussian components in multipath reflection, are merely examples. Those skilled in the art can adjust these parameters based on actual sensor performance, bridge structure complexity, and computational resources; this embodiment does not impose any limitations on these adjustments.

[0075] This embodiment achieves, for the first time, systematic modeling and parameterization of optical transmission characteristics, multipath reflection, and polarization characteristics in an under-bridge inspection scenario by collaboratively processing data from quantum lidar and visual cameras. By constructing an enhanced feature map that integrates geometric and physical effects, it provides a robust and information-rich matching foundation for adaptive positioning in GPS-denied environments. Through a data-driven polarization contamination decoupling network, it effectively distinguishes between inherent surface properties and environmental interference, improving the reliability of feature extraction. This embodiment achieves effective coupling of multiple physical effect parameters and enhanced map representation, improving the adaptability and positioning accuracy of the UAV perception system in extreme environments, and enhancing the accuracy and environmental adaptability of subsequent blind-zone positioning and matching.

[0076] Furthermore, this embodiment provides a method for acquiring real-time GPS navigation signal strength and calculating the distance between the UAV and the blind zone boundary, performing environmental prediction based on a bridge structure model and a physical effect propagation model before entering the blind zone, obtaining a pre-loaded physical effect compensation parameter set based on the environmental prediction results, and initializing a dual SLAM fusion positioning system based on quantum lidar and a visual sensor. The method includes:

[0077] Based on real-time GPS navigation signal strength, the distance between the UAV and the blind zone boundary is obtained by comparing signal strength thresholds and using a distance calculation model;

[0078] Based on the pre-set bridge structure model and real-time environmental data, the illumination distribution, multipath reflection area and polarization change trend in the blind zone are predicted by the under-bridge ray tracing model and physical effect propagation model, and the environmental prediction results are obtained.

[0079] Based on the environmental prediction results, a pre-loaded set of physical effect compensation parameters is obtained by matching and retrieving from the physical effect coupling database. The pre-loaded set of physical effect compensation parameters includes optical transfer function compensation parameters, polarization characteristic compensation parameters, and multipath reflection suppression parameters.

[0080] Based on the preloaded physical effect compensation parameter set and the local feature map under the bridge, an initialized dual SLAM fusion localization system is obtained through parallel startup and parameter loading processes. The dual SLAM fusion localization system includes a laser synchronous localization and mapping system and a visual synchronous localization and mapping system.

[0081] The signal strength threshold comparison involves comparing the real-time GPS navigation signal strength with a preset signal strength threshold to determine if the GPS signal is about to fail, providing a basis for switching positioning modes. The distance calculation model is a mathematical model that combines real-time UAV positioning data with a bridge structure model to calculate the straight-line distance from the UAV's current position to the blind zone boundary, quantifying the proximity of the UAV to the blind zone. This model is derived by incorporating bridge structure constraint factors based on the Euclidean distance formula. Real-time environmental data consists of environmental parameters affecting the propagation of physical effects collected in real-time during UAV flight, including real-time meteorological data, ambient light intensity, and atmospheric visibility, acquired through airborne meteorological and light sensors. The under-bridge ray tracing model is an improved ray propagation simulation model used to simulate the propagation of light in the complex structure beneath the bridge. The reflection and attenuation processes are analyzed to predict the illumination distribution and multipath reflection areas within the blind zone. This is achieved by adding atmospheric scattering correction terms and structural occlusion judgment logic to the traditional ray tracing algorithm. The environmental prediction results are quantitative predictions of physical effects within the blind zone based on the bridge structure model and real-time environmental data. These include the illumination distribution, multipath reflection areas, and polarization variation trends within the blind zone, providing a basis for preloading compensation parameters. The illumination distribution within the blind zone consists of the illumination intensity and direction distribution data at different spatial locations within the blind zone, obtained through simulation using a ray tracing model under the bridge. The multipath reflection areas are a set of spatial regions within the blind zone where laser light is prone to multiple reflections, identified by identifying the intersection areas of reflection paths using a ray tracing model. The polarization variation trend represents the variation of the polarization angle and degree of polarization of light at different locations within the blind zone with spatial location, predicted through a physical effect propagation model.

[0082] The pre-loaded physical effect compensation parameter set is a combination of parameters pre-loaded before entering the blind zone to counteract environmental physical effect interference. It includes optical transfer function compensation parameters, polarization characteristic compensation parameters, and multipath reflection suppression parameters, obtained through feature matching retrieval of environmental prediction results and a physical effect coupling database. The optical transfer function compensation parameters are used to correct ambient light transmission interference, obtained through optical transfer function model inversion, and are used for inverse filtering compensation of the image. The polarization characteristic compensation parameters are used to enhance the effectiveness of polarization information in visual features, obtained based on polarization characteristic analysis results, and are used to improve the decoupling effect between materials and contamination. The multipath reflection suppression parameters are used to filter false echoes from multipath reflections in laser point clouds, obtained through multipath reflection path analysis, and are used for net... The system utilizes quantum lidar point cloud data. The initialized dual-SLAM fusion localization system is a parallel localization system based on quantum lidar and a visual sensor. It includes a laser simultaneous localization and mapping (SLAM) system and a visual simultaneous localization and mapping (SLAM) system, obtained through a parallel startup algorithm and parameter loading process, and is used for fusion localization within blind zones. The laser simultaneous localization and mapping system is a localization and mapping system based on quantum lidar data, initialized after loading a geometric feature and physical effect coupling database, and used to provide geometric localization information. The visual simultaneous localization and mapping system is a localization and mapping system based on visual sensor data, initialized after loading visual features and a dynamic optical transfer function model, and used for auxiliary localization and feature verification.

[0083] Specifically, such as Figure 3 As shown, the first step is to analyze the GPS navigation signal strength, i.e., the carrier-to-noise ratio, collected in real time by the airborne GPS receiver module. Perform a signal strength threshold comparison and then... With the preset signal strength threshold Real-time comparison is performed; simultaneously, a distance calculation model is initiated, which is based on the Euclidean distance formula and incorporates bridge structural constraint factors. ,in This is a coefficient related to the degree of shading by the bridge structure, with a value range of [value missing]. The improved distance calculation formula is obtained. ,in The real-time location coordinates of the drone are obtained by fusing GPS positioning results with IMU data. The coordinates of the blind zone boundary feature points extracted from the bridge structural model are used to calculate the distance between the UAV and the blind zone boundary. Then, based on the preset bridge structure model, namely the pre-stored high-precision bridge CAD model and real-time environmental data, the under-bridge ray tracing model is launched. The under-bridge ray tracing model adds an atmospheric scattering correction term to the traditional ray tracing algorithm. and ,in For wind speed, For humidity, , Based on real-time environmental data, and by tracing the propagation path from the laser emission point to the bridge structure surface, the light intensity attenuation and reflection probability of each path are calculated to predict the light intensity distribution matrix within the blind zone. With the set of coordinates of the multipath reflection region , The number of multipath reflection regions is determined; simultaneously, a physical effect propagation model is initiated, employing an improved Long Short-Term Memory (LSTM) network. Its input consists of real-time environmental data and local bridge structural features, while its output is the polarization angle variation curve. With polarization degree change curve During network training, the number of hidden layers was set to 3, the number of neurons in each layer to 64, the number of iterations to 800, and the learning rate to 0.0015. The network was trained using a historical dataset of environmental physical effects monitoring under bridges. Real-time data was input into the model to obtain the polarization change trend.

[0084] Subsequently, based on the environmental prediction results, including the illumination distribution matrix, multipath reflection regions, and polarization change curves, the K-nearest neighbor feature matching algorithm is used to retrieve the pre-loaded physical effect compensation parameter set from the physical effect coupling database. The feature vectors from the environmental prediction results, including the mean illumination intensity, multipath reflection region density, and polarization change rate, are then compared with the feature vectors of the parameter set stored in the database. A similarity threshold is set. And select those with similarity greater than or equal to the similarity threshold. The parameter set serves as a pre-loaded set of physical effect compensation parameters, including optical transfer function compensation parameters, polarization characteristic compensation parameters, and multipath reflection suppression parameters. Finally, the laser SLAM system and the visual SLAM system are started simultaneously using a parallel startup algorithm. The laser SLAM system loads geometric feature data and multiple physical effect parameters extracted from the physical effect coupling database, and initializes the point cloud registration module and map update thread. The visual SLAM system loads the visual feature library and optical transfer function model, and initializes the image feature extraction module and polarization information processing thread. At the same time, the pre-loaded physical effect compensation parameter set and the local feature map under the bridge are loaded into the fusion layer of the dual SLAM system through the parameter interface, and the physical effect consistency monitoring thread is started to complete the initialization of the dual SLAM fusion positioning system.

[0085] For example, this embodiment assumes that the bridge to be inspected is a prestressed concrete T-beam bridge, and the coordinates of the feature points at the blind zone boundary in the bridge structural model are: The preset signal strength threshold is Bridge structural constraint factors Similarity threshold The quantum lidar scanning frequency is 12Hz, the binocular vision camera frame rate is 32fps, and the real-time environmental data is wind speed. ,humidity Ambient light intensity First, the real-time GPS signal strength was acquired through the GPS receiver module. With respect to the preset signal strength threshold Comparison shows that the real-time GPS signal strength meets the following requirements. Signal strength less than the preset threshold The real-time positioning coordinates of the drone are Substituting into the distance calculation formula, we get Then, the atmospheric scattering correction term was calculated using a ray tracing model under the bridge. and The weights for wind speed and humidity are respectively... , The average illumination intensity in the core region of the predicted blind zone is 500 lux. The multipath reflection region contains the coordinates of 12 feature points. The physical effect propagation model outputs the polarization angle variation curve. exist hour , hour ,in The polarization angle is a function of time, with units of . The curve representing the change in the polarization direction of light with the flight time of the UAV is shown. exist hour , hour ,in It is a function of polarization degree changing with time, and its value range is... The polarization purity of light is characterized by its value; the larger the value, the more pronounced the polarization characteristics.

[0086] Next, the environmental prediction feature vector is extracted and matched with the parameter set feature vector in the physical effect coupling database. A parameter set with a similarity of 0.88 is retrieved, resulting in the preloaded physical effect compensation parameter set. ,in These are the optical transfer function compensation parameters. 0.72 corresponds to the atmospheric attenuation compensation coefficient, used to counteract the attenuation effect of the atmosphere on light transmission. 0.68 is the scattering compensation coefficient, used to correct image blurring caused by environmental scattering. These are polarization characteristic compensation parameters. 32° corresponds to the polarization angle calibration value, used to correct polarization measurement deviations. 0.33 is the polarization degree enhancement coefficient, used to improve the recognition of polarization information. The parameters are multipath reflection suppression parameters. 0.15 corresponds to the multipath reflection energy suppression threshold, used to filter multipath reflection echoes with energy below this threshold. 2.2ns is the multipath reflection time threshold, used to eliminate spurious echo signals with flight times exceeding this threshold. Finally, the laser SLAM and visual SLAM systems are started in parallel, and the laser SLAM loads geometric feature points. For each physical effect parameter, the visual SLAM system loads a visual feature library containing 5000 feature templates and an OTF model, loads a pre-loaded compensation parameter set and a local feature map under the bridge, starts a physical effect consistency monitoring thread, and completes the initialization of the dual SLAM fusion positioning system. Further exemplarily, when the real-time GPS signal strength is... Signal strength greater than the preset threshold , Greater than the preset distance threshold When GPS remains the primary positioning source, the preloading process is not initiated; when real-time environmental data is... At that time, atmospheric scattering correction term The physical effect propagation model predicts an increase in the rate of change of polarization angle. The multipath reflection suppression parameters obtained from the search are: In this embodiment, the preset signal strength threshold, preset distance threshold, similarity threshold, and number of LSTM network iterations are merely examples. Those skilled in the art can set these values ​​according to the actual GPS signal environment, bridge structure size, and computing resources. This embodiment does not impose any restrictions on these values.

[0087] This embodiment achieves accurate determination of the boundary between the UAV and the blind zone by comparing signal strength thresholds and using an improved distance calculation model. It achieves accurate prediction of environmental physical effects within the blind zone by using a bridge-based ray tracing model with added environmental correction terms and an improved LSTM physical effect propagation model. It achieves rapid retrieval of compensation parameter sets by using a feature matching algorithm. It achieves efficient initialization of the dual SLAM system by using a parallel startup and parameter loading process. It realizes the preloading of physical effect compensation parameters and the early readiness of the positioning system before entering the blind zone, thereby improving the positioning response speed, environmental adaptability, and positioning accuracy stability of the UAV after entering the GPS denied blind zone.

[0088] Furthermore, this embodiment provides a method for compensating for physical effects in GPS-denied blind zones by using quantum lidar as an active optical probe in collaboration with a visual sensor. It also involves performing branch decisions based on the laser scanning matching effect, and using a covariance reconstruction algorithm with decoupled physical effects and a multi-physical constraint factor graph optimization method to perform fusion positioning calculations to obtain the optimized pose estimate, physical reliability grid, and real-time effect compensation degree for the UAV. The steps include:

[0089] A quantum lidar is used as an active optical probe to emit coded light pulses, and a visual sensor is used to acquire synchronously acquired images. Based on the coded light pulses and synchronously acquired images, a physical effect compensation is performed through an optical transfer function collaborative inversion algorithm to obtain the real-time optical transfer function. Inverse filtering compensation is performed on the synchronously acquired images to obtain optically compensated synchronously acquired images.

[0090] Based on the synchronously acquired images after optical compensation, the polarization information is extracted, and the pollution level estimation and material classification results are obtained through the polarization characteristics and pollution decoupling model.

[0091] Real-time evaluation of the matching confidence between LiDAR point cloud data and the local feature map under the bridge; execution of branch decisions based on the matching confidence:

[0092] When the matching confidence level is greater than or equal to a preset matching confidence threshold, the laser-synchronized localization and mapping (LSRT) system is used as the primary localization source, with visual LSRT as the auxiliary verification system, to obtain a first temporary localization result, and / or...

[0093] When the matching confidence is less than the preset matching confidence threshold, the weight of the visual synchronous localization and mapping system is increased, and the polarization and material decoupling model is activated to enhance visual features, thereby obtaining a second temporary localization result.

[0094] Based on the first or second temporary positioning result, combined with the pollution level estimation and material classification result, the optimized pose estimation, physical confidence grid and real-time effect compensation degree are obtained by performing fusion positioning calculation through the covariance reconstruction algorithm of physical effect decoupling and the multi-physical constraint factor graph optimization method.

[0095] The system comprises several key components: coded light pulses (CKPs) are laser pulses with time-coded information emitted by the quantum lidar, obtained by modulating their amplitude or frequency; synchronously acquired images are images of the bridge substructure simultaneously acquired by a visual sensor (i.e., a binocular vision camera) and the quantum lidar CKPs, with a sensor time synchronization mechanism ensuring consistent acquisition timing; a collaborative optical transfer function (OPF) inversion algorithm is an algorithm based on collaborative data inversion of the ambient light transmission characteristics using quantum lidar CKPs and synchronously acquired images, used to eliminate interference from physical effects such as uneven illumination and scattering; a real-time optical transfer function (RTF) is a dynamic function describing the light transmission attenuation and scattering characteristics in the current under-bridge environment, used to guide inverse image filtering compensation; and inverse filtering compensation is based on real-time optical transfer function (RTF)... The transfer function performs inverse filtering on the synchronously acquired images to counteract image blurring and color distortion caused by environmental physical effects. The synchronously acquired images after optical compensation are clear images that have undergone inverse filtering compensation, eliminating most of the interference from environmental physical effects, providing high-quality data for polarization information extraction and feature matching. The polarization characteristic and pollution decoupling model is a model used to separate the inherent polarization characteristics of the object surface from the additional polarization effect of the pollution layer, used to achieve material classification and pollution degree quantification, and is constructed based on an improved convolutional neural network. The pollution degree estimation is a quantitative value characterizing the severity of pollution on the bridge structure surface, with a value range of [0,1]. The material classification result is the identification result of the material on the bridge structure surface, such as concrete, metal, paint, etc.

[0096] The matching confidence score is the feature similarity score between the real-time point cloud data of the quantum lidar and the local feature map under the bridge, calculated by the iterative nearest point algorithm, and used to evaluate the reliability of laser SLAM positioning. The preset matching confidence threshold is the critical value for judging the matching effect of laser scanning, and is calibrated through a large number of experiments under the bridge. The first temporary positioning result is the UAV pose estimation obtained when laser SLAM is the main method and visual SLAM is the auxiliary method. The polarization and material decoupling model is a model used to enhance the discriminative power of visual features. By separating the inherent properties of the surface and the interference of the contamination layer, it improves the positioning accuracy of visual SLAM in laser feature degradation scenarios. It is trained based on polarization information and local geometric features of lidar. The second temporary positioning result is the UAV pose estimation obtained after the visual SLAM weight is increased, which is suitable for laser point cloud feature degradation scenarios.

[0097] Specifically, such as Figure 4 As shown, the quantum lidar first emits coded light pulses, simultaneously triggering the binocular vision camera to acquire images. The frame rate is consistent with the lidar's scanning frequency. Based on the emission timing of the coded light pulses, the time-of-flight histogram data, and the pixel intensity distribution of the simultaneously acquired images, the optical transfer function cooperative inversion algorithm is initiated. This algorithm, based on the traditional light transmission equation, introduces a time decay factor for the quantum lidar's coded pulses. ,in The time decay coefficient is obtained by linearly fitting the quantum lidar echo time histogram with the synchronous image light intensity data, and characterizes the time decay rate of light in the environment under the bridge. To determine the laser's time of flight, an improved transmission equation is constructed. ,in, To encode the incident light intensity of the optical pulse. For image receiving light intensity, The wavelength of light For pixel coordinates, For noise, To determine the camera response efficiency, the real-time optical transfer function is obtained by solving the equation using the gradient descent method. Based on real-time optical transfer function Inverse filtering compensation is performed on the synchronously acquired images. Using the Wiener filtering algorithm, the ratio of the noise power spectrum to the signal power spectrum is set to [value missing]. The synchronously acquired image after optical compensation is obtained.

[0098] Then, for the optically compensated synchronously acquired images, feature regions such as the bridge structure edges and bolts are selected. These feature regions are areas with obvious geometric or polarization characteristics, such as the bridge structure edges and bolts. The number of feature regions is set according to the image complexity. The Stokes vector of each feature region is then calculated. The input polarization characteristics and contamination decoupling model is an improved CNN architecture containing three convolutional layers with 3×3 kernels, and 16, 32, and 64 kernels respectively. It also includes two pooling layers (max pooling with a stride of 2) and two fully connected layers. The model takes Stokes vectors and the local point cloud normal vectors from the LiDAR as inputs to obtain a contamination level estimate. Based on the material classification probability distribution, the training process uses the Adam optimizer with an initial learning rate of 0.002, 600 iterations, and a loss function that is a weighted sum of cross-entropy and mean squared error with a weight ratio of 3:7. The dataset was trained using a bridge-under-the-bridge scene dataset labeled with pollution levels and materials. Then, the iterative nearest-point algorithm was used to calculate the matching confidence between the real-time LiDAR point cloud and the local feature map under the bridge. , and the preset matching confidence threshold Perform a comparison; if a match is found, the confidence level is... Greater than or equal to the preset matching confidence threshold Then the laser SLAM weights are set to Visual SLAM weights are set to And satisfy The laser SLAM output pose is verified and corrected by visual SLAM to obtain the first temporary localization result. If the confidence level matches Less than the preset matching confidence threshold Then, by activating the polarization and material decoupling model, the visual SLAM weight is increased and the laser SLAM weight is decreased. The second temporary localization result is obtained through image texture enhancement and marker extraction. .

[0099] The final result was based on the first provisional localization. Or the second provisional location result Using the initial value, combined with the pollution level estimate Based on the material classification results, a covariance reconstruction algorithm decoupled from physical effects is initiated to decompose the sensor noise into inherent noise. Pollution level estimation Positively correlated environmental coupling noise Reconstruct the covariance matrix ,in The coupling coefficient is used; an optical transmission consistency factor and a material and polarization correlation factor are introduced into the factor diagram, and the optimized pose estimate is obtained by optimization using the Gauss-Newton method. Based on optimized pose estimation The covariance matrix is ​​used to construct a physical reliability grid using a three-dimensional gridded mapping and uncertainty propagation method. The real-time effect compensation degree is obtained by weighted fusion of pollution degree estimation and optical transfer function inversion error. .

[0100] For example, this embodiment assumes that the bridge to be inspected is a steel-concrete composite beam bridge, and the quantum lidar scanning frequency is [missing information]. Binocular vision camera frame rate The preset matching confidence threshold is The preset matching confidence threshold is the critical threshold for judging the matching effect of laser scanning. It is obtained by collecting matching data of laser point clouds and feature maps of 1000 different under-bridge scenes, and then calibrating the minimum similarity required to achieve the positioning accuracy. First, the quantum lidar emits coded light pulses, and simultaneously acquires images with a resolution of 1920×1080. The time decay coefficient of the real-time optical transfer function is obtained through collaborative inversion of the optical transfer function. After inverse filtering compensation, the optically compensated synchronously acquired image is obtained; 10 feature regions are selected from the optically compensated synchronously acquired image, and the Stokes vector of a certain bolt region is calculated as follows: ,in For three-dimensional Stokes vectors, This represents the total light intensity grayscale value of the area. This represents the grayscale difference between the 0° and 90° polarization directions. The grayscale difference between the 45° and 135° polarization directions is used as input. Through a decoupling model of polarization characteristics and pollution, the pollution level is estimated as follows: The material classification result is metal. The material classification result is obtained by matching the model with the similarity between the inherent polarization characteristics and the preset material polarization template. The polarization degree of metal materials is usually lower than that of materials such as concrete.

[0101] Then, the mean squared error after point cloud registration is calculated using the iterative nearest point algorithm, and the matching confidence is obtained. Less than the preset matching confidence threshold Activate the polarization and material decoupling model to improve the visual SLAM weights and obtain the visual SLAM weights as follows: The second temporary location result was obtained. ,in The x-axis coordinate of the UAV in the local navigation coordinate system represents the horizontal position of the UAV in the direction of bridge extension, calculated based on the geometric reference of the local feature map under the bridge. The y-axis coordinate of the UAV in the local navigation coordinate system represents its horizontal position perpendicular to the direction of bridge extension. Similarly, it is calculated based on the map reference. Let z be the z-axis coordinate of the UAV in the local navigation coordinate system, representing its height relative to the bridge deck reference plane. The roll angle is the angle of rotation of the drone around its own x-axis, which is obtained from the visual SLAM attitude calculation. The pitch angle of the drone is the rotation angle of the drone around its own y-axis, calculated using the same method as visual SLAM. The yaw angle of the UAV is the rotation angle of the UAV around its own z-axis, used to characterize the UAV's heading; the second temporary positioning result is used. Using the initial values, the covariance matrix is ​​reconstructed and multi-physics constraint factor graph optimization is introduced to obtain the optimized pose estimate. .when Under conditions of relatively low pollution, with a matching confidence level of Greater than the preset matching confidence threshold At that time, the laser SLAM weights were set to The first temporary location result was obtained. Similarly, the first temporary location result Using the initial values, the covariance matrix is ​​reconstructed and multi-physics constraint factor graph optimization is introduced to obtain the optimized pose estimate. The matching confidence threshold, visual SLAM weights, and time decay coefficients preset in this embodiment are merely examples. Those skilled in the art can adjust them according to the actual sensor performance and bridge environment. This embodiment does not impose any limitations on these adjustments.

[0102] This embodiment achieves improved robustness and accuracy of UAV pose estimation in GPS-denied environments under bridges by compensating for the synergistic physical effects of quantum lidar and visual sensors, adapting to different positioning scenarios through a branch decision mechanism based on matching confidence, separating noise interference through a covariance reconstruction algorithm with decoupled physical effects, and realizing high-precision positioning through multi-physical constraint factor graph optimization. It solves the problem of positioning failure in scenarios with degraded laser features and enhances the environmental adaptability of quantum navigation and quantum positioning.

[0103] Furthermore, this embodiment provides a step for obtaining enhanced navigation and positioning results by updating the physical effect coupling database through dual-path closed-loop calibration based on structural features and dynamic perturbations, based on optimized pose estimation, including:

[0104] Based on optimized pose estimation, when the UAV flies close to the pre-stored key structural features, the positioning drift is corrected by structural feature calibration through dual-path closed-loop calibration based on the local feature map under the bridge, and the first calibration result is obtained.

[0105] Based on optimized pose estimation, when airflow disturbances are detected causing flight attitude fluctuations, the positioning deviation is corrected by dynamic disturbance calibration through dual-path closed-loop calibration based on inertial measurement unit data and visual image data from multi-source sensor navigation data, and a second calibration result is obtained.

[0106] Based on the first and second calibration results, the optical transfer function model and the correlation model between polarization characteristics and material properties are refitted and cross-validated online to obtain the cross-validation results.

[0107] The physical effects coupling database is updated based on the cross-validation results, and the optimized pose estimation is recalculated based on the updated physical effects coupling database to obtain the enhanced navigation and positioning results.

[0108] Furthermore, this embodiment provides a step of updating the physical effect coupling database based on cross-validation results, and recalculating the optimized pose estimation based on the updated physical effect coupling database to obtain enhanced navigation and positioning results, including:

[0109] Based on the cross-validation results, the parameter update amount of the validated model is obtained through parameter increment calculation and validity verification;

[0110] Based on the validated model parameter update, the corresponding optical transfer function parameters, polarization characteristic parameters, and multipath reflection characteristic parameters in the physical effect coupling database are updated through database indexing and replacement operations.

[0111] Based on the updated physical effect coupling database, the enhanced navigation and positioning results are obtained by reloading the updated model parameters at most physical constraint factor graph optimization framework and optimizing and recalculating the optimized pose estimation.

[0112] The process involves several key structural features: Pre-stored key structural features are typical structural elements with precise geometric coordinates extracted from the bridge structure model and stored in the local feature map under the bridge, including bolt hole centers, weld endpoints, and support anchor points. These features are obtained by importing the bridge CAD model and executing a feature extraction algorithm. Structural feature calibration is a process of correcting the UAV positioning result using the deviation between the actual coordinates of the pre-stored key structural features and the current detection coordinates, eliminating long-term accumulated positioning drift. Positioning drift is a systematic error caused by the accumulation of sensor noise and the time-varying environmental physical effects, leading to a gradual deviation of the UAV pose estimation from the true value over time. The first calibration result is the corrected UAV pose estimation output after structural feature calibration, with reduced drift error. Dynamic disturbance calibration is a process of detecting instantaneous changes in UAV attitude caused by airflow disturbances based on inertial measurement unit data and visual optical flow data, and compensating for the positioning estimate in real time, suppressing positioning deviations caused by environmental dynamic disturbances. Positioning deviation is a random error caused by external dynamic disturbances, resulting in an instantaneous deviation of the UAV pose estimation from the true value. The second calibration result is the real-time corrected UAV pose estimation output after dynamic disturbance calibration, with the deviation caused by the disturbance suppressed.

[0113] The optical transfer function model is a dynamic mathematical model describing the light distribution, scattering, and color distortion effects during light transmission in the environment under the bridge. It is obtained through optical transmission characteristic modeling and is used to model and compensate for ambient light transmission interference. The polarization characteristic and material property correlation model is a statistical or neural network model describing the mapping relationship between the polarization reflection characteristics of different material surfaces under specific lighting conditions and their material categories. It is obtained through polarization characteristic analysis and material classification training and is used to achieve material identification and decoupling from contamination. Online refitting and cross-validation involves re-estimating the parameters of the optical transfer function model and the polarization characteristic and material property correlation model using new observation data corresponding to the first and second calibration results. The effectiveness is verified by comparing the consistency of the output results of the two models, and this is used to realize physical effect modeling. The system employs several mechanisms: online adaptive updates; cross-validation results serve as a consistency measure and parameter validity indicator between the outputs of the two models after online refitting, guiding the updating of the physical effect coupling database; parameter increment calculation and validity verification calculate the change in model parameters relative to the original database values ​​based on the cross-validation results, and determine the reliability and usability of the increment through a preset consistency index threshold, ensuring the safety and accuracy of database updates; validated model parameter update amounts are the set of parameter changes that have passed validity verification and can be used to update the physical effect coupling database; and database indexing and replacement operations locate the corresponding entry in the physical effect coupling database based on the spatial location index of the parameter, replacing the original parameter value with the validated parameter update amount, enabling real-time incremental updates of the database.

[0114] Specifically, such as Figure 5 As shown, when the drone flies close to the pre-stored key structural features, the three-dimensional coordinates of the feature are retrieved from the local feature map under the bridge. Simultaneously, the observation coordinates of this feature in the current frame are detected in real time using lidar and visual sensors. Calculate coordinate deviation An outlier removal method based on Mahalanobis distance was adopted, with a removal bias greater than [value missing]. Unreliable observations, among which The standard deviation of the uncertainty in structural feature localization is obtained through historical observation statistics; then, the optimal pose estimation is used. Using these as initial values, construct structural feature constraint factors and add them to the multiphysics constraint factor graph. The structural feature constraint factors are defined as follows: ,in From position The observation model to the characteristic observation coordinates, The corrected pose is obtained by optimizing the observation noise covariance matrix of the structural features using the Gauss-Newton method. This is the first calibration result; subsequently, when airflow disturbance is detected, the abrupt changes in angular velocity and acceleration are extracted from the inertial measurement unit data. Simultaneously, the optical flow field is calculated from the visual image sequence, and its divergence and vorticity characteristics are extracted. When the amplitude of the angular velocity abrupt change exceeds a preset abrupt change amplitude threshold... Or the optical flow divergence exceeds the preset optical flow divergence threshold. When the event occurs, it is considered a valid disturbance event.

[0115] Based on the inertial measurement unit data and visual optical flow data within the time period corresponding to the effective disturbance event, a dynamic disturbance observation model is constructed. ,in Angular velocity, For acceleration, For the optical flow field, an extended Kalman filter framework is employed to optimize pose estimation. Given the initial state, the perturbation-compensated pose is obtained through state prediction and observation updates. That is, the second calibration result; the final result is based on the first calibration result. Compared with the second calibration result Optical and polarization observation data for the corresponding time periods were extracted separately, and the optical transfer function model was refitted online using a recursive least squares algorithm with a forgetting factor. The forgetting factor was set to 1. Update the optical transfer function parameter vector The correlation model between polarization characteristics and material properties was refitted online using the online gradient descent method, with a learning rate set to [value missing]. Update the model weights; perform cross-validation on the outputs of the two models on the same observation data, and calculate their output consistency index. ,in and These are the predicted outputs of the two models for the same observation, respectively. If the consistency index... Greater than or equal to the preset consistency index threshold Then the cross-validation is deemed successful, and the cross-validation result is obtained. Based on cross-validation results The parameter entries corresponding to spatial locations in the physical effect coupling database are indexed and replaced using a B+ tree index structure with spatial hash values ​​as keys. and Update to the database; finally, reload the updated parameter-maximum physical constraint factor graph optimization framework to optimize pose estimation. A new round of optimization was performed, resulting in enhanced navigation and positioning results. .

[0116] For example, this embodiment assumes the substructure of a steel box girder bridge, and the bridge span... The drone is currently located near a pre-stored bolt feature. Its pre-stored coordinates are The observed coordinates were detected. The standard deviation of the structural feature positioning uncertainty is The calculated coordinate deviation is After outlier removal, the results are retained; structural feature constraint factors are constructed, and the first calibration result is obtained through factor graph optimization. Simultaneously, an airflow disturbance event was detected, causing a sudden change in the amplitude of the angular velocity of the inertial measurement unit. Exceeding the preset mutation amplitude threshold , optical divergence Exceeding the preset optical flow divergence threshold The disturbance was determined to be valid; a second calibration result was obtained based on the extended Kalman filter. .

[0117] Subsequently, optical and polarization observation data for the corresponding time periods were extracted, and the optical transfer function model was refitted online to update the parameter vector. The correlation model between polarization characteristics and material properties is updated online using gradient descent to obtain new weights. ; Calculate the consistency index of the outputs of the two models And the consistency index is greater than the preset consistency index threshold. Cross-validation passed; based on the successful cross-validation, the corresponding entries in the physical effect coupling database were updated and reloaded into the factor graph optimization framework to optimize the original pose estimation. Optimization and recalculation are performed to obtain enhanced navigation and positioning results. ;when and At that time, structural feature calibration and dynamic disturbance calibration are still performed, but the correction amount is smaller, and the model update magnitude is correspondingly reduced, ultimately... and The differences are subtle. In this embodiment, the standard deviation of structural feature positioning uncertainty, the preset abrupt change amplitude threshold, the preset optical flow divergence threshold, and the preset consistency index threshold are merely examples. Those skilled in the art can set these based on the actual bridge structural feature accuracy, airflow disturbance intensity, and model output stability requirements; this embodiment does not impose any limitations on these settings.

[0118] This embodiment employs a dual-path closed-loop mechanism of structural feature calibration and dynamic disturbance calibration to correct long-term cumulative drift and instantaneous disturbance deviations, respectively. It achieves adaptive updates of the physical effect model through online model refitting based on recursive least squares and online gradient descent. Cross-validation ensures the consistency and reliability of model updates. Real-time incremental updates of the physical effect coupled database are achieved through database indexing and replacement operations, ultimately yielding enhanced navigation and positioning results. This enables the UAV to continuously maintain and adaptively improve its positioning accuracy in GPS-denied environments under bridges, thereby enhancing the environmental adaptability and operational reliability of inspection tasks.

[0119] Furthermore, this embodiment provides a step for enhancing navigation and positioning results to generate flight control commands for a UAV, including:

[0120] Based on the physical reliability grid and real-time effect compensation degree, the physical stability evaluation value of the current region is obtained through the physical stability assessment model.

[0121] Based on the physical stability evaluation value, enhanced navigation and positioning results, and real-time effect compensation degree, a physical fusion predictive control model is constructed and configured.

[0122] Flight control commands are obtained based on the configured physical fusion predictive control model and the UAV dynamics model.

[0123] Among them, physical stability assessment is an evaluation process based on physical reliability grid and real-time effect compensation degree to calculate the comprehensive stability level of the current flight area, used to quantify the friendliness of the environment to the navigation system; physical fusion predictive control model is a model predictive control model that incorporates physical stability evaluation value as a penalty term into the cost function, used to adaptively adjust flight trajectory and control commands according to environmental perception quality; UAV dynamics model is a mathematical model describing the mapping relationship between the force and motion state of the UAV in three-dimensional space, obtained through mechanism modeling combined with experimental identification methods, used to accurately predict the UAV's response behavior to control commands, providing a constraint basis for solving the control model; flight control commands are control signal sequences obtained by solving the rolling time domain optimization problem through physical fusion predictive control model, used to directly drive the UAV's actuators; the physical stability evaluation value of the current area is a scalar value that quantifies the current flight area environment's support capability for navigation and positioning accuracy and flight control safety, with a value range of [missing information]. The larger the value, the more stable the environment and the higher the navigation reliability. The global average reliability and real-time effect compensation degree based on the physical reliability grid are calculated by a weighted summation formula and are used to dynamically adjust the conservatism of the flight control strategy.

[0124] Specifically, the physical stability evaluation value of the current region is first calculated, and then the current position of the UAV and the preset future flight path are traversed in the physical reliability grid. Extract the confidence level of each voxel from all voxels covered within the inner layer. Calculate the global average confidence level ,in The total number of voxels covered by the path; based on the real-time effect compensation degree. A physical stability assessment model is constructed, and the physical stability evaluation value of the current region is obtained through a weighted fusion formula. ,in The weighting coefficients are satisfied. The calibration was obtained through offline simulation experiments. , To ensure that the assessment results simultaneously reflect both positioning reliability and environmental compensation effectiveness; if the current area's physical stability evaluation value Less than the preset physical stability safety threshold If the current area is marked as a high-risk environment, a conservative adjustment of the control strategy is triggered. Then, a physical fusion predictive control model is constructed and configured. This model is an improvement on the model predictive control framework, introducing a physical stability evaluation value as a dynamic penalty term. Its core structure includes a prediction time domain... Control Time Domain Cost function and constraints; prediction time domain Its corresponding time length , To control the period, control the time domain. That is, the output of each control cycle Each control variable is executed only once, and then the optimization is performed continuously.

[0125] The cost function is designed as follows ,in For trajectory tracking error term, ,in , For reference position and velocity, To predict position and velocity, This is the weight matrix. For the smoothing term of the control quantity ,in To control the quantity, For a smooth weight matrix, This is the physical stability penalty coefficient. Enhancing penalties in high-risk environments to prevent aggressive control; constraints include UAV dynamic constraints derived from UAV dynamics models and control quantity saturation constraints, such as motor speed range. Obstacle avoidance constraints are obtained by setting no-fly zones in low-confidence areas within a physical confidence grid; model training involves offline collection of 1000 sets of flight data under different environments on a bridge, and the cost function is optimized using an improved interior-point method, with an iterative convergence threshold set at [value missing]. The maximum number of iterations is This ensures the real-time performance and accuracy of the control command solution; then, the enhanced navigation and positioning results are loaded. Extract current position Location of the target inspection point Generate reference trajectory ( The reference trajectory is generated using cubic polynomial interpolation to ensure continuous and smooth position and velocity; the physical stability evaluation value is used. Real-time effect compensation degree Reference trajectory and the current status of the drone, including its location. ,speed ,attitude Input the configured physical fusion predictive control model; finally, combine the UAV dynamics model to solve the control commands, embedding the UAV dynamics model as a constraint into the optimization problem of the physical fusion predictive control model. The state equation of the dynamics model is: , where the state vector Control vector This corresponds to the rotational speeds of the four motors in a quadcopter drone; the optimal control sequence is obtained by solving the rolling time-domain optimization of model predictive control within each control cycle. Take the first control quantity Flight control commands for the current cycle, including motor speed and attitude adjustment, directly drive the drone's actuators.

[0126] For example, this embodiment assumes that the bridge to be inspected is a prestressed concrete continuous beam bridge, and the drone is currently performing a bolt inspection task. The enhanced navigation and positioning result is... The target inspection point is The preset future flight path length in the physical credibility grid is... The total number of voxels covered is The calculated global average confidence level is The real-time effect compensation degree is Substituting into the physical stability assessment formula, the physical stability evaluation value of the current region is obtained. Physical fusion predictive control model settings The weight matrix is , , Penalty coefficient The reference trajectory is generated through cubic polynomial interpolation, and the spacing between adjacent reference points is... Input the above data into the model and solve it using the UAV dynamics model, where the inertia matrix is... Aerodynamic coefficient The current cycle flight control command is obtained as follows That is, the rotational speed of the four motors of the drone and the attitude adjustment amount. The drone maintains stable flight as it approaches the target point; when the global average confidence level is... Real-time effect compensation degree At that time, the calculated evaluation value of physical stability was... It is below the preset physical stability safety threshold. The model will automatically increase the penalty coefficient. By tightening the control variable variation constraints, a conservative control command is obtained. Flight speed from Down to Control cycle maintained To ensure flight safety; in this embodiment The values ​​are merely examples, and those skilled in the art can set them according to the actual performance of the UAV, the complexity of the bridge environment, and the requirements of the inspection task. This embodiment does not impose any limitations on this.

[0127] This embodiment quantifies the environment's support capability for flight control through a physical stability assessment model, integrates multi-dimensional perception information through an improved physical fusion predictive control model, provides precise response constraints through a UAV dynamics model, and achieves real-time optimal control through rolling time-domain optimization. This enables adaptive adjustment of UAV flight control in GPS-denied environments, improving flight stability, target alignment accuracy, and operational safety in complex under-bridge environments.

[0128] Furthermore, this application provides an adaptive flight control system for unmanned aerial vehicles (UAVs) in a GPS-denied environment under bridge substructure. This system includes a data acquisition and mapping module, a system preheating module, a positioning compensation module, and a closed-loop calibration module.

[0129] Mapping module: Acquires navigation data from multiple sources of the UAV, including at least lidar point cloud data acquired by quantum lidar, visual image data acquired by visual sensors, and inertial measurement unit (IMU) data acquired by IMU. Based on the navigation data from multiple sources of the UAV, a physical effect coupling database and a local feature map of the environment under the bridge are established through optical transmission characteristic modeling, polarization characteristic analysis, and multipath reflection path analysis.

[0130] System preheating module: acquires real-time GPS navigation signal strength and calculates the distance between the UAV and the blind zone boundary. Before entering the blind zone, it performs environmental prediction based on the bridge structure model and the physical effect propagation model. Based on the environmental prediction results, it obtains the preloaded physical effect compensation parameter set and initializes the dual SLAM fusion positioning system with quantum lidar and visual sensor as the core.

[0131] Positioning compensation module: In the blind zone where GPS is denied, the quantum lidar is used as an active optical probe to perform physical effect compensation in collaboration with the visual sensor. Based on the laser scanning matching effect, branch decision is executed. Through the covariance reconstruction algorithm of physical effect decoupling and the multi-physical constraint factor graph optimization method, the fusion positioning solution is performed to obtain the optimized pose estimation, physical credibility grid and real-time effect compensation degree of the UAV.

[0132] Closed-loop calibration module: Based on optimized pose estimation, it updates the physical effect coupling database through dual-path closed-loop calibration based on structural features and dynamic disturbances, and obtains enhanced navigation and positioning results.

[0133] The system preheating module, positioning compensation module, and closed-loop calibration module are all located on the server. The server receives raw sensor data transmitted from the acquisition devices via a wireless communication link and performs real-time analysis. The acquisition devices include a quantum lidar mounted on the UAV, a binocular vision camera with a global shutter, and an inertial measurement unit. The acquisition and mapping module can be deployed on the server or the UAV's onboard computing unit. After completing initial data acquisition and map construction within the GPS effective area, it uploads the physical effect coupling database and the local feature map under the bridge to the server. The modules are connected and exchange information through data interfaces. The acquisition and mapping module provides the physical effect coupling database and the local feature map under the bridge to the system preheating module; the system preheating module provides the positioning compensation module with a pre-loaded physical effect compensation parameter set and an initialized dual SLAM system; the positioning compensation module provides the closed-loop calibration module with optimized pose estimation, physical reliability grid, and real-time effect compensation degree; the enhanced navigation and positioning results output by the closed-loop calibration module are finally sent to the UAV's flight controller to generate adaptive flight control commands.

[0134] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0135] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0136] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0137] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An adaptive flight control method for unmanned aerial vehicles (UAVs) in a GPS-denied environment under bridge substructure, characterized in that, include: We acquire navigation data from multiple sources of UAV sensors, and obtain a physical effect coupling database and a local feature map under the bridge by modeling optical transmission characteristics, analyzing polarization characteristics and resolving multipath reflection paths. By acquiring real-time GPS navigation signal strength and combining it with local feature maps under the bridge, an initial dual-SLAM fusion positioning system is obtained through environmental prediction using a bridge structure model and a physical effect propagation model. Based on the initialization dual SLAM fusion localization system, the optimal pose estimate of the UAV is obtained through physical effect compensation, covariance reconstruction algorithm and multi-physical constraint factor graph optimization method. Based on optimized pose estimation, the physical effect coupling database is updated through dual-path closed-loop calibration to obtain enhanced navigation and positioning results.

2. The adaptive flight control method for UAVs under GPS denial conditions in bridge substructure according to claim 1, characterized in that, Acquire navigation data from multiple sources of the UAV's sensors, and obtain a physical effect coupling database and a local feature map under the bridge through optical transmission characteristic modeling, polarization characteristic analysis, and multipath reflection path resolution; including: Based on navigation data from multiple sources, optical transmission function parameters are obtained by modeling optical transmission characteristics, and multipath reflection characteristic parameters are obtained by analyzing multipath reflection paths. Polarization characteristic parameters are obtained based on multi-source sensor navigation data through Stokes vector calculation and polarization characteristic analysis. Based on optical transfer function parameters, polarization characteristic parameters, and multipath reflection characteristic parameters, a physical effect coupled database and a local feature map under the bridge are constructed.

3. The adaptive flight control method for UAVs under GPS denial conditions in bridge substructure according to claim 1, characterized in that, By acquiring real-time GPS navigation signal strength and combining it with a local feature map under the bridge, an initial dual-SLAM fusion positioning system is obtained through environmental prediction using a bridge structure model and a physical effect propagation model. This system includes: Based on real-time GPS navigation signal strength, the distance between the UAV and the blind zone boundary is obtained by comparing signal strength thresholds and using a distance calculation model; Based on the pre-set bridge structure model and real-time environmental data, environmental prediction results are obtained through the under-bridge ray tracing model and the physical effect propagation model. Based on the environmental prediction results, a preloaded set of physical effect compensation parameters is obtained by matching and retrieving from the physical effect coupling database. Based on the preloaded physical effect compensation parameter set and the local feature map under the bridge, an initialized dual SLAM fusion positioning system is obtained through parallel startup and parameter loading processes.

4. The adaptive flight control method for unmanned aerial vehicles (UAVs) under GPS denial conditions in bridge substructure as described in claim 1, characterized in that, Based on the initial dual-SLAM fusion localization system, an optimized pose estimate for the UAV is obtained through physical effect compensation, covariance reconstruction algorithm, and multi-physics constraint factor graph optimization method, including: Based on quantum lidar and vision sensor, the optically compensated synchronously acquired image is obtained by performing physical effect compensation and inverse filtering compensation through optical transfer function cooperative inversion algorithm. Based on synchronously acquired images after optical compensation, pollution level estimation and material classification results are obtained through a decoupling model of polarization characteristics and pollution. The matching confidence of the lidar point cloud data and the local feature map under the bridge is evaluated in real time, and a branch decision is performed based on the matching confidence to obtain a temporary localization result; Based on the temporary positioning results, the optimized pose estimation, physical credibility grid, and real-time effect compensation degree are obtained by using the covariance reconstruction algorithm and the multi-physics constraint factor graph optimization method.

5. The adaptive flight control method for unmanned aerial vehicles (UAVs) under GPS denial conditions in bridge substructure as described in claim 4, characterized in that, Based on the matching confidence, branch decisions are executed to obtain provisional localization results, including: When the matching confidence level is greater than or equal to a preset matching confidence threshold, a first temporary positioning result is obtained by using the laser-synchronized localization and mapping system as the primary positioning source and the visual-synchronized localization and mapping system as the auxiliary verification, and / or, When the matching confidence is less than the preset matching confidence threshold, the weight of the visual synchronous localization and mapping system is increased, and the polarization and material decoupling model is activated to enhance visual features, thus obtaining a second temporary localization result.

6. The adaptive flight control method for unmanned aerial vehicles (UAVs) under GPS denial conditions in bridge substructure as described in claim 4, characterized in that, Based on quantum lidar and a vision sensor, a synchronously acquired image after optical compensation is obtained through physical effect compensation and inverse filtering compensation using an optical transfer function cooperative inversion algorithm, including: The quantum lidar is used as an active optical probe to emit coded light pulses, and the visual sensor is used in conjunction to acquire synchronously collected images. Based on the coded light pulses and synchronously acquired images, the light transmission equation in the under-bridge environment is solved and inverted using an optical transfer function collaborative inversion algorithm to obtain the real-time optical transfer function; Based on the real-time optical transfer function, the synchronously acquired image is subjected to inverse filtering to obtain an optically compensated synchronously acquired image.

7. The adaptive flight control method for unmanned aerial vehicles (UAVs) under GPS denial conditions in bridge substructure as described in claim 4, characterized in that, Based on synchronously acquired images with optical compensation, pollution level estimation and material classification results are obtained through a decoupling model of polarization characteristics and pollution, including: Polarization information is obtained by calculating the Stokes vector of the feature region based on the synchronously acquired image after optical compensation. Based on multi-source sensor navigation data, the polarization information is decomposed into inherent polarization characteristics and additional polarization effects through a polarization characteristic and pollution decoupling model, and the decomposition result is obtained. Based on the decomposition results, the pollution level is estimated by quantifying the additional polarization effect, and the material classification result is obtained by analyzing and judging the inherent polarization characteristics.

8. The adaptive flight control method for unmanned aerial vehicles (UAVs) under GPS denial conditions in bridge substructure as described in claim 4, characterized in that, Based on the provisional localization results, optimized pose estimation, physical reliability grid, and real-time effect compensation are obtained through a covariance reconstruction algorithm and a multi-physics constraint factor graph optimization method, including: Based on the pollution level estimation and material classification results, the decoupled sensor observation noise covariance matrix is ​​calculated using a covariance reconstruction algorithm that decouples physical effects. Using the temporary positioning result as the initial value, and combining the decoupled sensor observation noise covariance matrix, the optimized pose estimate is obtained through the multi-physics constraint factor graph optimization method. Based on the covariance of the optimized pose estimation, a physical credibility grid is obtained through three-dimensional mesh mapping and uncertainty propagation method; Based on the pollution level estimation and optical transfer function collaborative inversion algorithm, the real-time effect compensation degree is obtained through weighted fusion.

9. The adaptive flight control method for unmanned aerial vehicles (UAVs) under GPS denial conditions in bridge substructure according to claim 1, characterized in that, Based on optimized pose estimation, the physical effect coupling database is updated through dual-path closed-loop calibration to obtain enhanced navigation and positioning results, including: Based on optimized pose estimation, when the UAV flies close to the pre-stored key structural features, the positioning drift is corrected by structural feature calibration through dual-path closed-loop calibration and the first calibration result is obtained. Based on optimized pose estimation, when airflow disturbances are detected that cause flight attitude fluctuations, the positioning deviation is corrected by dynamic disturbance calibration through dual-path closed-loop calibration and a second calibration result is obtained. Based on the first and second calibration results, the optical transfer function model and the correlation model between polarization characteristics and material properties are refitted and cross-validated online to obtain the cross-validation results. The physical effect coupling database is updated based on the cross-validation results, and the optimized pose estimation is recalculated based on the updated physical effect coupling database to obtain the enhanced navigation and positioning results.

10. The adaptive flight control method for unmanned aerial vehicles (UAVs) under GPS denial conditions in bridge substructure according to claim 1, characterized in that, The multi-source sensor navigation data includes at least lidar point cloud data acquired by quantum lidar, visual image data acquired by visual sensors, and inertial measurement unit data acquired by inertial measurement units.

Images