Unmanned Aerial Vehicle (UAV) Autonomous Inspection System Based on Historical Inspection Visual Features

Abstract

This invention relates to the field of unmanned aerial vehicle (UAV) control and discloses a signalless autonomous inspection system for UAVs based on historical inspection visual features. The system includes: a historical decoupled feature storage module, a signal integrity monitoring module, a dynamic parameter sensing module, a field-of-view coordinate system decoupled transformation module, and a pose correction and compensation module. The transformation module constructs a transformation matrix based on real-time attitude angle data and encoder values, mapping the field-of-view sampled feature points to an absolute horizontal reference plane, generating a feature set that removes projection distortion caused by aerodynamic attitude deflection due to the aircraft's resistance to airflow. The compensation module generates a pose compensation term based on the topological residuals of the feature elements in the set and storage module, driving the aircraft along the inspection trajectory. This invention achieves physical decoupling between visual matching and aerodynamic attitude, eliminating positioning drift caused by differences in environmental wind fields.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle control technology, and in particular relates to a signal-free autonomous inspection system for unmanned aerial vehicles based on historical inspection visual features. Background Technology

[0002] Currently, in situations where global satellite navigation system signals are lost, the industry's conventional practice is to rely on pre-stored offline maps or to construct high-precision 3D models through active flight training. Real-time images captured by airborne cameras are then used to perform feature matching with prior data to determine spatial attitude. As an unsteady flight platform, rotorcraft are constrained by aerodynamic laws. When passing through the same spatial coordinate point at different times, in order to resist the randomly changing environmental wind field and maintain thrust balance, the airframe will produce different pitch and roll attitude deflections. This physical deviation in attitude directly leads to nonlinear perspective distortion in the imaging perspective of the camera components, introducing huge geometric errors into the 2D feature point matching logic at the underlying level.

[0003] To correct perspective deviations, existing improvement approaches typically focus on increasing the complexity of image enhancement algorithms or employing deep learning models for semantic correction. However, this does not physically eliminate the interference path of wind field disturbances on visual navigation systems. In addition to hardware mechanical stability limitations, related control methods face bottlenecks when dealing with the dynamic decoupling of aerodynamic attitude and visual features. For example, Chinese invention patent application CN118052743A discloses an image distortion correction method for UAV multi-attitude photography. By acquiring camera intrinsic parameters and calculating the coordinate transformation matrix, it performs image correction based on weak perspective projection relationships. The premise for the weak perspective projection model is that the depth change of the target object is much smaller than the object distance. In real inspection environments, such as encountering deep valley elevation differences or severe wind fields causing high-frequency aerodynamic attitude deflection, this projection model suffers from a fundamental deviation due to the mismatch of objective physical boundaries. It cannot achieve real-time alignment of visual features and aerodynamic response physical quantities in three-dimensional space, resulting in positioning accuracy and response frequency failing to meet the closed-loop inspection requirements under extreme conditions.

[0004] Therefore, the technical problem to be solved by this invention is how to utilize the visual features recorded during the historical inspection of rotorcraft, combined with a real-time aerodynamic attitude decoupling mechanism, to achieve reliable navigation, positioning and inspection task closed loop under no-signal conditions. Summary of the Invention

[0005] This invention aims to solve the problem of visual navigation failure and inspection mission interruption caused by the deflection of the aircraft's attitude due to environmental wind fields during signalless inspection of rotorcraft.

[0006] In this technical solution, a signal-free autonomous inspection system for unmanned aerial vehicles (UAVs) based on historical inspection visual features is provided. The system includes:

[0007] The historical decoupling feature storage module is used to incrementally store the visual decoupling feature elements, three-dimensional spatial coordinates, and task payload control command sequences of key locations in the inspection trajectory.

[0008] The signal integrity monitoring module is used to acquire the strength value of external navigation signals in real time, and generate an autonomous inspection trigger signal when the strength value is lower than the preset navigation loss threshold for a preset duration threshold.

[0009] The dynamic parameter sensing module is used to collect real-time attitude angle data of the flight platform body module and encoder values ​​of the load stability control module.

[0010] The field of view coordinate system decoupling transformation module is used to construct a three-dimensional spatial transformation matrix based on real-time attitude angle data and encoder values, and to inversely project the feature points in the field of view characterization sampling data onto the absolute horizontal reference plane defined by the inertial state perception module based on the three-dimensional spatial transformation matrix, thereby generating a two-dimensional horizontal projection feature set that is stripped of the projection distortion caused by the aerodynamic attitude deflection caused by the flight platform body module resisting the wind field airflow.

[0011] The pose correction and compensation module is used to respond to the autonomous inspection trigger signal, calculate the topological geometric residual between the two-dimensional horizontal projection feature set and the visual decoupling feature elements of the corresponding key position points in the historical decoupling feature storage module, and drive the flight control module based on the pose compensation term generated by the topological geometric residual, so that the flight platform body module moves along the inspection trajectory. When the pose compensation term indicates that the flight platform body module has reached the key position point, the module issues a sequence of mission payload control commands.

[0012] Preferably, the field of view coordinate system decoupling transformation module follows the following logic during projection transformation: construct a rotation compensation matrix based on the roll and pitch angles in the real-time attitude angle data, and determine the relative pose relationship between the sampling field of view center axis and the flight platform body module based on the encoder values; combine the rotation compensation matrix and the relative pose relationship to complete the three-dimensional rotation inverse transformation of the feature points in the field of view characterization sampling data, and map them to the absolute horizontal reference plane.

[0013] Preferably, the pose correction and compensation module includes a multi-dimensional state estimation sub-module. The multi-dimensional state estimation sub-module is used to use the generated positioning result as a position offset compensation term, and uses the Kalman filter algorithm to fuse the position offset compensation term with the acceleration quadratic integral displacement output by the inertial state perception module to correct the cumulative drift deviation of the flight platform body module in the absence of a signal.

[0014] Preferably, the historical decoupling feature storage module adopts an incremental index structure, and after the autonomous inspection trigger signal is generated, the feature index logic of the corresponding trajectory segment is activated based on the last known pose before the signal is lost.

[0015] Preferably, the task load control command sequence includes sampling switch trigger parameters and load pointing angle parameters. The pose correction and compensation module is used to send the corresponding sampling switch trigger parameters to the task sensor module when the pose compensation item meets the preset action trigger threshold.

[0016] Preferably, the dynamic parameter sensing module is used to acquire pitch angle data, roll angle data, and heading angle data of the flight platform body module at a sampling frequency higher than 50Hz.

[0017] Preferably, the multidimensional state estimation submodule is also used to monitor the intensity of changes in ambient irradiance, and when the intensity of changes in ambient irradiance exceeds a preset threshold of 1000 lx, causing abnormalities in the normalized matching residuals, the measurement noise term corresponding to the positioning result in the Kalman filter algorithm is increased.

[0018] Preferably, the visual decoupling feature elements in the historical decoupling feature storage module are data elements stored after the three-dimensional rotation inverse transformation is completed on the absolute horizontal reference plane by the field of view coordinate system decoupling transformation module during the preset teaching and inspection phase.

[0019] Preferably, the system also includes a redundant safety control module, which is used to drive the power drive control module to switch the flight platform body module to the stationary vertical landing mode when the signal integrity monitoring module determines that the external navigation signal has been continuously lost for more than a preset 300s time limit.

[0020] Compared with existing technologies, the UAV signal-free autonomous inspection system based on historical inspection visual features has the following advantages:

[0021] 1. In unmanned aerial vehicle (UAV) autonomous inspection without signal, this invention achieves physical decoupling between the visual matching process and the aerodynamic wind-resistant attitude of the rotorcraft, eliminating positioning drift caused by differences in the environmental wind field. This invention utilizes the attitude parameters output by the airborne inertial measurement unit and the joint angles fed back by the gimbal joint encoder to construct a spatial rigid body transformation relationship. The real-time acquired two-dimensional feature points are forcibly projected onto a virtual horizontal reference plane with the body's center of mass as the origin and perpendicular to the gravity vector. Topological matching is then performed on this reference plane with the historical feature library. This mechanism transforms the complex three-dimensional perspective deviation into a deterministic geometric compensation process, eliminating the interference of aerodynamic attitude changes on visual positioning accuracy from the physical source, and ensuring that the system can obtain a stable attitude reference under different wind field conditions.

[0022] 2. Constructing an incremental passive feature acquisition and task sequence triggering linkage mode to ensure the logical continuity of inspection operations in the absence of signals. During the normal inspection phase, this invention uses a silent background method to record the image feature descriptors and spatial coordinates of each inspection node, eliminating the need for a dedicated flight training modeling process. When the multi-source radio signal monitoring unit determines that the external navigation signal is lost, the system automatically activates the feature matching logic of the corresponding flight line segment based on the index relationship of the current pose in the historical feature library, and simultaneously retrieves the camera shutter trigger command and gimbal pointing angle associated with that position. This processing method, which deeply binds the navigation pose with the specific operation task, enables the aircraft not only to complete the flight path following, but also to autonomously complete the detection work of the remaining nodes according to the original inspection sequence and original shooting requirements, realizing the transformation from simple flight line blind spot filling to full mission closed-loop operation.

[0023] 3. Improve the processing efficiency of airborne computing resources under emergency conditions and enhance the system's response speed. Because this invention completes the coordinate transformation based on the horizontal reference plane during the historical feature storage stage, and only incrementally stores the topological coordinates and descriptors of key frame feature points, rather than storing the entire original image data, it significantly reduces the storage and retrieval load of the airborne processor. In the signalless autonomous inspection mode, the processor only needs to perform geometric comparison of the feature set in the two-dimensional horizontal projection space stripped of attitude parameters, avoiding the use of high-power image affine transformation or deep learning perspective correction operators. This technical path based on physical kinematic parameter pre-correction shortens the logical link of single-frame pose calculation and ensures that the refresh frequency of control commands meets the dynamic response requirements of aircraft in complex space environments. Attached Figure Description

[0024] Figure 1 This is a diagram showing the core functional modules and multi-source data interaction architecture of the system of this invention;

[0025] Figure 2 This is a diagram showing the state switching and logic flow of the entire autonomous inspection process of this invention. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0027] It should be noted that all directional and positional terms used in this invention, such as: up, down, left, right, front, back, vertical, horizontal, inner, outer, top, bottom, transverse, longitudinal, center, etc., are only used to explain the relative positional relationship and connection between components in a specific state (as shown in the accompanying drawings). They are only for the convenience of describing this invention and do not require that this invention be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention. In addition, the descriptions of "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0028] In the description of this invention, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections; they can refer to direct connections or indirect connections through an intermediate medium; they can refer to the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0029] In the description of this specification, references to the terms "an embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example, and the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0030] A signal-free autonomous inspection system for unmanned aerial vehicles (UAVs) based on historical inspection visual features, the system comprising:

[0031] The historical decoupling feature storage module is used to incrementally store the visual decoupling feature elements, three-dimensional spatial coordinates, and task payload control command sequences of key locations in the inspection trajectory.

[0032] The signal integrity monitoring module is used to acquire the strength value of external navigation signals in real time, and generate an autonomous inspection trigger signal when the strength value is lower than the preset navigation loss threshold for a preset duration threshold.

[0033] The dynamic parameter sensing module is used to collect real-time attitude angle data of the flight platform body module and encoder values ​​of the load stability control module.

[0034] The field of view coordinate system decoupling transformation module is used to construct a three-dimensional spatial transformation matrix based on real-time attitude angle data and encoder values, and to inversely project the feature points in the field of view characterization sampling data onto the absolute horizontal reference plane defined by the inertial state perception module based on the three-dimensional spatial transformation matrix, thereby generating a two-dimensional horizontal projection feature set that is stripped of the projection distortion caused by the aerodynamic attitude deflection caused by the flight platform body module resisting the wind field airflow.

[0035] The pose correction and compensation module is used to respond to the autonomous inspection trigger signal, calculate the topological geometric residual between the two-dimensional horizontal projection feature set and the visual decoupling feature elements of the corresponding key position points in the historical decoupling feature storage module, and drive the flight control module based on the pose compensation term generated by the topological geometric residual, so that the flight platform body module moves along the inspection trajectory. When the pose compensation term indicates that the flight platform body module has reached the key position point, the module issues a sequence of mission payload control commands.

[0036] Preferably, the field of view coordinate system decoupling transformation module follows the following logic during projection transformation: construct a rotation compensation matrix based on the roll and pitch angles in the real-time attitude angle data, and determine the relative pose relationship between the sampling field of view center axis and the flight platform body module based on the encoder values; combine the rotation compensation matrix and the relative pose relationship to complete the three-dimensional rotation inverse transformation of the feature points in the field of view characterization sampling data, and map them to the absolute horizontal reference plane.

[0037] Preferably, the pose correction and compensation module includes a multi-dimensional state estimation sub-module. The multi-dimensional state estimation sub-module is used to use the generated positioning result as a position offset compensation term, and uses the Kalman filter algorithm to fuse the position offset compensation term with the acceleration quadratic integral displacement output by the inertial state perception module to correct the cumulative drift deviation of the flight platform body module in the absence of a signal.

[0038] Preferably, the historical decoupling feature storage module adopts an incremental index structure, and after the autonomous inspection trigger signal is generated, the feature index logic of the corresponding trajectory segment is activated based on the last known pose before the signal is lost.

[0039] Preferably, the task load control command sequence includes sampling switch trigger parameters and load pointing angle parameters. The pose correction and compensation module is used to send the corresponding sampling switch trigger parameters to the task sensor module when the pose compensation item meets the preset action trigger threshold.

[0040] Preferably, the dynamic parameter sensing module is used to acquire pitch angle data, roll angle data, and heading angle data of the flight platform body module at a sampling frequency higher than 50Hz.

[0041] Preferably, the multidimensional state estimation submodule is also used to monitor the intensity of changes in ambient irradiance, and when the intensity of changes in ambient irradiance exceeds a preset threshold of 1000 lx, causing abnormalities in the normalized matching residuals, the measurement noise term corresponding to the positioning result in the Kalman filter algorithm is increased.

[0042] Preferably, the visual decoupling feature elements in the historical decoupling feature storage module are data elements stored after the three-dimensional rotation inverse transformation is completed on the absolute horizontal reference plane by the field of view coordinate system decoupling transformation module during the preset teaching and inspection phase.

[0043] Preferably, the system also includes a redundant safety control module, which is used to drive the power drive control module to switch the flight platform body module to the stationary vertical landing mode when the signal integrity monitoring module determines that the external navigation signal has been continuously lost for more than a preset 300s time limit.

[0044] Example 1: In a power transmission line inspection mission deployed in a high-altitude mountainous environment, the UAV faces a situation where the global navigation system signal is physically interrupted because the inspection route traverses deep valleys. When the signal integrity monitoring module determines that the strength of the external navigation signal is lower than the preset navigation loss threshold for a certain duration, the system generates an autonomous inspection trigger signal. During operation in this signal-free environment, the UAV encounters a crosswind. To maintain the preset spatial tracking position and thrust balance, the flight platform module adjusts the pitch angle. With roll angle Generates aerodynamic resistance, among which, The pitch angle represents the angle at which the aircraft rotates about its horizontal axis. This represents the roll angle of the aircraft rotating around its longitudinal axis. The aerodynamic attitude deflection of the aircraft causes nonlinear perspective distortion in the camera components mounted on the load stabilization control module, resulting in a mismatch between the real-time acquired 2D image features and the visual decoupling feature elements in the historical decoupling feature storage module. The dynamic parameter sensing module acquires the aircraft's pitch angle, roll angle, and yaw angle data in real time, and simultaneously obtains the encoder values ​​fed back by the gimbal motor encoder, transmitting the physical parameters to the field-of-view coordinate system decoupling transformation module. The field-of-view coordinate system decoupling transformation module constructs a 3D spatial transformation matrix based on the real-time attitude angle data, and inversely projects the 2D feature points in the field-of-view characterization sampling data onto the [other surface] based on this 3D spatial transformation matrix. On the absolute horizontal reference surface defined by the inertial state perception module, a set of two-dimensional horizontal projection features is generated, stripped of the projection distortion caused by the aerodynamic attitude deflection induced by the flight platform body module resisting the wind field airflow. This stripping of projection distortion is based on the physical foundation of the far-field coplanar feature assumption. Since the physical distance between the background target and the camera in the inspection environment is much greater than the local undulation height difference of the target itself, the retained feature points are approximately located in the same three-dimensional infinite plane in terms of geometric topology. Under this physical boundary condition, the three-dimensional perspective projection model of the target area naturally degenerates into a two-dimensional affine transformation model in terms of mathematical mechanism, thus making it physically feasible to directly and inversely cancel the three-dimensional perspective distortion through the two-dimensional projection matrix.

[0045] The pose correction and compensation module responds to the autonomous inspection trigger signal, calculates the topological geometric residual between the two-dimensional horizontal projection feature set and the visual decoupling feature elements of the incremental index in the historical decoupling feature storage module, and obtains the translation vector and yaw angle of the current aircraft relative to the historical inspection trajectory. The multi-dimensional state estimation submodule inside the system uses the Kalman filter algorithm to use the generated positioning result as the position offset compensation term, and fuses it with the acceleration quadratic integral displacement output by the inertial state perception module to correct the cumulative drift deviation of the flight platform body module in the absence of signal. When the multi-dimensional state estimation submodule detects that the intensity of the change in environmental irradiance exceeds the preset 1000lx threshold, normalization matching is triggered. When residual anomalies occur, the system increases the measurement noise term corresponding to the positioning result in the Kalman filter algorithm and switches to a mode where the heading is maintained by inertial navigation calculation. Here, lx represents lux, which is the unit of illumination. The pose correction and compensation module drives the flight control module based on the fused and corrected pose compensation term. Under conditions of no external radio navigation signal and fluctuating wind field, the flight platform module travels along the original inspection trajectory. When the pose compensation term indicates that the flight platform module has reached the preset key position point, the system issues a sequence of mission payload control commands to drive the mission sensor module to complete the detection work of the remaining inspection nodes according to the original shooting requirements, realizing the connection from the signal interruption point to the full mission closed-loop inspection state.

[0046] In the infrared temperature measurement of insulators in substations, the system maintains the calling logic of the sequence of control commands for the task load. When the deviation magnitude between the real-time spatial position calculated by the pose correction and compensation module and the spatial coordinates of the preset inspection node in the historical decoupling feature storage module is within 0.2m, the system determines that the UAV has entered the task triggering area and extracts the shooting angle and exposure parameters corresponding to the node from the historical decoupling feature storage module. It then drives the task sensor module to execute the shutter command. By mapping and associating the physical operation of the task load with the geometric coordinates based on the absolute horizontal reference plane, the UAV can sequentially reproduce the inspection and detection actions of the teaching phase in an environment without external navigation signals. Based on the projection mechanism of the absolute horizontal reference plane of the airframe aerodynamic attitude decoupling, the three-dimensional field of view deviation compensation process is transformed into a deterministic spatial geometric cancellation method based on the UAV's native inertia and joint feedback hardware. This cuts off the transmission path of environmental wind field interference to the visual navigation system, and the system obtains a fixed pose reference under different wind conditions.

[0047] Example 2: On a physical simulation platform simulating airflow fluctuations in high-altitude mountainous areas, the system operates under conditions where the ambient wind speed continuously varies from 0 m / s to 15 m / s. The sampling frequency of the dynamic parameter sensing module is set to 100 Hz. This sampling frequency setting is intended to balance the filtering effect of the high-frequency mechanical vibration signals generated by the flight platform module with the real-time performance of attitude calculation. If the sampling frequency is lower than 50 Hz, it will be unable to capture transient attitude jitter caused by crosswinds, resulting in a lag bias in the transformation matrix generated by the field-of-view coordinate system decoupling transformation module. If the sampling frequency is higher than 200 Hz, unnecessary data redundancy will occur, increasing the computational load on the airborne processor.

[0048] In the standard operating condition test with an ambient wind speed of 5.2 m / s, to counteract the effects of crosswinds, the pitch angle in the real-time attitude angle data generated by the flight platform's main module... The roll angle is 8.15°. It is -4.23°, where The pitch angle represents the angle at which the aircraft rotates about its horizontal axis. This represents the roll angle of the aircraft rotating around its longitudinal axis; at this time, the encoder value fed back by the load stabilization control module indicates the angular offset of the camera component relative to the aircraft coordinate system; as a control group, without calling the field of view coordinate system decoupling transformation module, feature matching is directly performed on the field of view characterization sampling data, and the measured average value of the feature matching residual is 24.8 pixels, and the cumulative drift amount generated by the position offset compensation term output by the pose correction compensation module within 10 seconds reaches 1.86m; in the sample group of this invention, the field of view coordinate system decoupling transformation module is based on and A three-dimensional spatial transformation matrix M is constructed to map two-dimensional feature points to an absolute horizontal reference plane, generating a two-dimensional horizontal projection feature set. M is a 4×4 affine transformation matrix calculated based on the real-time attitude angle of the aircraft and the numerical solution of the gimbal encoder. At this time, the mean value of the feature matching residual is reduced to 2.1 pixels, and the cumulative drift is reduced to 0.12m. The experiment introduces 20dB Gaussian white noise interference generated by environmental obstructions to simulate the industrial inspection environment with changing light and shadow. When the multi-dimensional state estimation submodule detects that the environmental irradiance changes abruptly from 1200lx to 350lx, where lx is lux and represents the unit of illuminance, the normalized matching residual increases from 0.05 to 0.42. The system triggers a dynamic increase in the measurement noise term R in the Kalman filter algorithm, reducing the visual positioning weight to 15% of the initial value, where R represents the measurement noise covariance matrix in the Kalman filter. At this time, the system relies on the acceleration quadratic integral displacement output by the inertial state perception module to maintain the heading calculation.

[0049] By setting comparative samples with different wind speed gradients, it was found that when the ambient wind speed is between 0 m / s and 12 m / s, the positioning error output by the attitude correction and compensation module shows a linear low growth trend with the increase of wind speed, and the error increment remains at 0.02 m / m / s. When the ambient wind speed exceeds the performance inflection point of 12.5 m / s, the positioning error increases nonlinearly and dramatically, with the error increment jumping to 0.15 m / m / s, because the motor output power of the flight platform module reaches more than 95% of the rated upper limit and the attitude angle fluctuation frequency exceeds the dynamic response bandwidth of the field of view coordinate system decoupling and transformation module. This data verifies that the system has attitude adjustment stability in wind field environments within 12 m / s. Based on multi-dimensional gradient test data, the inspection system, under the condition of aerodynamic attitude deflection and illumination disturbance, suppresses the visual matching error to the pixel level through real-time decoupling of physical quantification parameters and multi-source data fusion compensation, confirming the deterministic nature of the solution for non-electric variable adjustment and the ability to maintain trajectory in signal-free environments.

[0050] Example 3: This example combines Figures 1 to 2 This paper describes a signalless autonomous inspection system for drones based on historical inspection visual features, such as... Figure 1As shown, the signal integrity monitoring module acquires the intensity value of the external navigation signal, generates an autonomous inspection trigger signal, and transmits the autonomous inspection trigger signal to the pose correction and compensation module. Simultaneously, the dynamic parameter sensing module collects the real-time attitude angle data of the flight platform body module and the encoder values ​​of the load stability control module, and transmits the real-time attitude angle data and encoder values ​​to the field of view coordinate system decoupling transformation module. The field of view coordinate system decoupling transformation module projects the data inversely onto the absolute horizontal reference plane to generate a two-dimensional horizontal projection feature set stripped of attitude deflection, and transmits it to the pose correction and compensation module. The historical decoupling feature storage module incrementally stores the visual decoupling feature elements and the mission load control command sequence, and transmits the visual decoupling feature elements and command sequence to the pose correction and compensation module. After receiving the above data, the pose correction and compensation module calculates the topological geometric residual and generates a pose compensation term to issue a mission load control command sequence. The multi-dimensional state estimation sub-module contained in the pose correction and compensation module uses Kalman filtering to fuse the position offset compensation term to correct the cumulative drift deviation without signal.

[0051] like Figure 2 As shown, the state transition mechanism of the system during operation is as follows: Figure 2 As shown, under normal inspection state, if the external navigation signal strength is lower than the threshold, the system state transitions to autonomous inspection trigger state. If the signal is lost for more than 300 seconds, the system state transitions to vertical autonomous landing state. After entering the autonomous inspection trigger state, the system activates the corresponding trajectory segment feature index and enters the field of view decoupling transformation state. It enters the pose fusion correction state by calculating the topological geometric residual. In the pose fusion correction state, if an abnormal environmental irradiance occurs, the system continues to cycle and maintain the pose fusion correction state. When the action trigger threshold is met, the system enters the mission action reproduction state from the pose fusion correction state. After executing the control command sequence in the mission action reproduction state, the system returns to the pose fusion correction state.

[0052] Example 4: During an internal inspection of a large-span steel structure factory building, the UAV enters a signal-shielded indoor space from an outdoor environment with satellite signals, and the system completes the navigation source switch. The navigation loss threshold in the signal integrity monitoring module is determined by the calibration procedure. The signal simulator is used to reduce the carrier-to-noise ratio of the global satellite navigation system signal in a controlled electromagnetic environment. The horizontal positioning accuracy factor of the positioning result is statistically analyzed. When the horizontal positioning accuracy factor exceeds the safety limit value of 3.5 and the positioning solution cannot be locked in a fixed solution for 2.0s, the median signal strength at this time is recorded as -145dBm. This value is set as the navigation loss threshold, and 2.0s is set as the preset duration threshold, where dBm is the unit of absolute power level.

[0053] When the signal integrity monitoring module detects that the external navigation signal strength is below -145dBm for more than 2.0 seconds, the system initiates autonomous inspection; the field-of-view coordinate system decoupling transformation module receives real-time attitude angle data and encoder values ​​transmitted by the dynamic parameter sensing module, and generates transformations from the field-of-view coordinate system to the body coordinate system and from the body coordinate system to the horizontal geographic coordinate system; the system then uses the pitch angle from the real-time attitude angle data... Roll angle And the pitch compensation angle of the gimbal motor A 3D spatial transformation matrix M is constructed by converting the pixel coordinates of real-time acquired feature points into vectors in the normalized camera coordinate system based on the camera intrinsic parameter matrix, and then multiplying them sequentially by... Define the gimbal decoupling matrix, by and Define the body attitude rotation matrix to counteract the perspective shift caused by the body tilt, and project it onto the absolute horizontal reference plane defined by the plane equation Z=H, where H is the real-time inspection height value.

[0054] In the event of loss of external navigation signals, the real-time inspection altitude value H is quantitatively obtained by a ranging sensor mounted on the bottom of the flight platform. Based on the physical principle of electromagnetic wave time-of-flight ranging, a single-line lidar or millimeter-wave radar with a sampling frequency of not less than 100Hz is used to emit a detection beam along the z-axis of the aircraft coordinate system and receive the ground echo to obtain the relative slant distance from the bottom surface of the aircraft to obstacles on the lower surface. Combined with the dynamic parameter sensing module, the pitch angle is output synchronously. With roll angle The data is used to perform geometric compensation perpendicular to the gravity vector for the relative slant distance through the trigonometric cosine function, filtering out the trigonometric elevation measurement error introduced by the aerodynamic attitude deflection of the aircraft, and outputting the compensated absolute ground clearance as the real-time inspection height value H for constructing the plane equation.

[0055] The currently generated 2D horizontal projection feature set is processed by the pose correction and compensation module. The multidimensional state estimation submodule uses a noise mapping function to adjust the parameters of the Kalman filter algorithm. When the data from the ambient irradiance sensor indicates that the rate of change of irradiance Δlx exceeds the slope threshold of 500lx / s within 0.5s, the system detects that the visual feature extraction quality is impaired. The pose correction and compensation module multiplies the measurement noise term R in the Kalman filter by a gain coefficient of 10.0, reducing the weight of the visual positioning result. Instead, the pose estimation is maintained by the integral of angular velocity and acceleration output by the inertial state perception module. During the transient process, the trajectory deviation is maintained to be less than 0.05m. After the visual environment stabilizes, the system recaptures the visual decoupling feature elements in the historical decoupling feature storage module based on the two-dimensional horizontal projection feature set, corrects the accumulated position residual, and enables the UAV to continuously move the mission payload along the inspection path under no signal and indoor light and shadow interference. The system transforms the visual tracking process into a geometric cancellation method driven by attitude sensor data by physically calibrating the navigation signal loss criterion and decoupling the field of view coordinate projection path, eliminating the discontinuity of visual matching under heterogeneous environmental conditions and maintaining the stability of the system's trajectory in the signal shielded area.

[0056] Example 5: During the initial equipment deployment in a new inspection area, the system establishes the physical reference for attitude calculation through a pre-set initial state calibration procedure. The dynamic parameter sensing module collects the acceleration vector G defined by the gravity field when the UAV is stationary. The field of view coordinate system decoupling transformation module determines the opposite direction of this vector G as the normal vector of the absolute horizontal reference plane. The system reads the encoder value at the zero position of the load stabilization control module, calculates the initial installation deviation angle δ between the imaging plane of the camera component and the absolute horizontal reference plane determined by the normal vector, and writes the static correction matrix constructed by the initial installation deviation angle δ into the storage unit of the airborne processor. This compensates for the physical installation zero point offset generated during the calculation of the three-dimensional spatial transformation matrix M of the subsequent real-time attitude angle data.

[0057] In ongoing power grid inspection operations, the system utilizes the feature solidification procedures in the historical decoupling feature storage module to improve the matching success rate in signal-free environments. During the teaching phase, the historical decoupling feature storage module acquires field-of-view characterization sampling data for different time periods. The field-of-view coordinate system decoupling transformation module projects the collected two-dimensional feature points onto the absolute horizontal reference plane based on real-time attitude angle data, generating feature space coordinate distributions corresponding to different observation times. The system uses a clustering method based on Euclidean spatial distance to identify visual decoupling feature elements whose position change frequency is lower than a preset frequency threshold in different inspection batches. It establishes an incremental index association between the topological coordinates of the visual decoupling feature elements and the task load control command sequence, and removes feature points with cluster center deviations greater than 0.05m, establishing the underlying reference coordinate data that supports the pose correction and compensation module in performing topological geometric residual calculations.

[0058] Example 6: In the complex shielded environment inside a substation, the UAV system initiates the inertial reference calibration program of the dynamic parameter sensing module before takeoff. When the flight platform body module is in a horizontal static state, the processor continuously collects 500 sets of raw data from the triaxial accelerometer, calculates the average projection of the gravity vector in the body coordinate system, and uses it as the initial gravity compensation operator to solidify into the multidimensional state estimation submodule. Based on this, the integral zero drift compensation value is determined by monitoring the discrete variance of the angular velocity in the static state. This compensation value is used to perform real-time zero-bias correction on the acceleration quadratic integral displacement during subsequent motion.

[0059] During the construction of the 3D spatial transformation matrix M by the field-of-view coordinate system decoupling transformation module, the system uses a projection method based on the pinhole camera model to process the field-of-view characterization sampling data. The field-of-view coordinate system decoupling transformation module reads the camera intrinsic parameter matrix K stored in non-volatile memory. This matrix K is a 3×3 matrix containing the focal length and principal point coordinates. The field-of-view coordinate system decoupling transformation module constructs a rotation operator by calculating the sine and cosine function values ​​of the pitch and roll angles in the real-time attitude angle data, and combines the real-time acquired feature point pixel coordinates with their corresponding real-time inspection height values ​​H, mapping the feature points to the absolute horizontal reference plane according to the following formula: ,in, Represents the coordinate vector of the absolute horizontal reference plane. This represents the total rotation matrix combining the aircraft's attitude and the gimbal offset. This represents a pixel coordinate vector containing real-time inspection height information. This procedure eliminates visual topological distortion caused by aircraft tilt under different wind field environments, ensuring that the system-generated 2D horizontal projection feature set is aligned with the coordinate data pre-stored in the historical decoupled feature storage module in Euclidean geometry. Addressing the inherent scale loss issue when mapping pixel coordinates to physical space, the system executes a defined depth compensation logic when constructing this pixel coordinate vector. It directly multiplies the homogeneous coordinates of the basic pixels on the 2D image plane by the real-time inspection height value fed back from the underlying ranging sensor and after triangulation compensation. This serves as a scale factor to transform the dimensionless ray vector into a 3D spatial coordinate vector with a true physical scale, thus ensuring the absolute validity of subsequent inverse matrix multiplication operations in Euclidean space. Before generating the 2D horizontal projection feature set, the field-of-view coordinate system decoupling transformation module introduces a parallax isolation procedure. Based on the physical laws of motion parallax, objects at different depths in 3D space... The displacement of the two-dimensional imaging plane projection is negatively correlated with the physical depth of field of the observation center. The field of view coordinate system decoupling transformation module extracts two consecutive frames of field of view characterization sampling data and calculates the optical flow divergence parameters of each feature point in the corresponding pixel space. In order to eliminate the scalar interference introduced by the absolute translation velocity of the UAV itself on the optical flow calculation, after extracting the original optical flow divergence value, the system extracts the magnitude of the real-time translation velocity vector output by the inertial state perception module as the normalization denominator and performs a division dimensionality reduction operation on the original optical flow divergence value. Then, the composite quantity affected by the coupling of the body velocity is transformed into a pure disparity characterization factor that is only absolutely related to the physical spatial depth. The optical flow divergence parameters are compared with the preset disparity divergence threshold set to three per thousand of the total pixel length of the diagonal of the sampled image. Near-field strong undulating structural feature points with optical flow divergence parameters greater than the disparity divergence threshold are filtered out, and far-field coplanar feature points with optical flow divergence parameters not greater than the threshold are retained and incorporated into the subsequent inverse projection transformation link to the absolute horizontal reference plane.

[0060] In autonomous fly-around inspections of power transmission towers, the attitude correction and compensation module executes a topological geometric residual calculation procedure. The system uses a nearest neighbor search method to filter matching feature point pairs in a two-dimensional horizontal projection feature set, and constructs a cost function for the translation vector and yaw correction angle based on the least squares criterion. By solving for the minimum value of this cost function, the attitude deviation of the current aircraft relative to the inspection trajectory is obtained. When the system determines that the confidence level of the matching feature points is greater than 0.85 and the spatial distribution uniformity meets the preset conditions, the historical decoupling feature storage module collects the new feature points after attitude decoupling and performs incremental storage. The coordinates of the new feature points are fused with the original feature map to achieve dynamic coverage of the inspection environment feature description as the environment changes. In addition, all the quantization thresholds used for system state switching and filtering logic were not subjectively set, but were obtained through offline calibration on a hardware-in-the-loop simulation platform that covers the gradual range of illumination change and multiple levels of wind speed and direction. The system traversed the body attitude angle offset in the wind tunnel environment and the adjustable light source test field, compared the visual topology deviation output with the absolute pose truth value recorded by the high-precision motion capture system in real time, and used the gradient descent method to find the parameter solution set that minimizes the cumulative pose drift. Thus, in engineering, the gain coefficient of 10.0 during illumination change, the empirical threshold of 0.3% for disparity divergence isolation, and the convergence critical value of 0.85 to ensure the lower limit of topology matching confidence were determined, ensuring that all system parameters have solid experimental physical boundary support.

[0061] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A signal-free autonomous inspection system for unmanned aerial vehicles (UAVs) based on historical inspection visual features, characterized in that, The system includes: The historical decoupling feature storage module is used to incrementally store the visual decoupling feature elements, three-dimensional spatial coordinates, and task payload control command sequences of key locations in the inspection trajectory. The signal integrity monitoring module is used to acquire the strength value of external navigation signals in real time, and generate an autonomous inspection trigger signal when the strength value is lower than the preset navigation loss threshold for a preset duration threshold. The dynamic parameter sensing module is used to collect real-time attitude angle data of the flight platform body module and encoder values ​​of the load stability control module. The field of view coordinate system decoupling transformation module is used to construct a three-dimensional spatial transformation matrix based on real-time attitude angle data and encoder values, and to inversely project the feature points in the field of view characterization sampling data onto the absolute horizontal reference plane defined by the inertial state perception module based on the three-dimensional spatial transformation matrix, thereby generating a two-dimensional horizontal projection feature set that is stripped of the projection distortion caused by the aerodynamic attitude deflection caused by the flight platform body module resisting the wind field airflow. The pose correction and compensation module is used to respond to the autonomous inspection trigger signal, calculate the topological geometric residual between the two-dimensional horizontal projection feature set and the visual decoupling feature elements of the corresponding key position points in the historical decoupling feature storage module, and drive the flight control module based on the pose compensation term generated by the topological geometric residual, so that the flight platform body module moves along the inspection trajectory. When the pose compensation term indicates that the flight platform body module has reached the key position point, the module issues a sequence of mission payload control commands.

2. The unmanned aerial vehicle (UAV) autonomous inspection system based on historical inspection visual features according to claim 1, characterized in that, The field of view coordinate system decoupling transformation module follows the following logic during projection transformation: construct a rotation compensation matrix based on the roll and pitch angles in the real-time attitude angle data, and determine the relative pose relationship between the sampling field of view center axis and the flight platform body module based on the encoder values; combine the rotation compensation matrix and the relative pose relationship to complete the three-dimensional rotation inverse transformation of the feature points in the field of view characterization sampling data, and map them to the absolute horizontal reference plane.

3. The unmanned aerial vehicle (UAV) autonomous inspection system based on historical inspection visual features according to claim 1, characterized in that, The pose correction and compensation module includes a multi-dimensional state estimation sub-module. The multi-dimensional state estimation sub-module is used to take the generated positioning result as the position offset compensation term and use the Kalman filter algorithm to fuse the position offset compensation term with the acceleration quadratic integral displacement output by the inertial state perception module to correct the cumulative drift deviation of the flight platform body module in the absence of signal.

4. The unmanned aerial vehicle (UAV) autonomous inspection system based on historical inspection visual features according to claim 1, characterized in that, The historical decoupling feature storage module adopts an incremental index structure, and after the autonomous inspection trigger signal is generated, it activates the feature index logic of the corresponding trajectory segment based on the last known pose before the signal is lost.

5. The unmanned aerial vehicle (UAV) autonomous inspection system based on historical inspection visual features according to claim 1, characterized in that, The mission load control command sequence includes sampling switch trigger parameters and load pointing angle parameters. The pose correction and compensation module is used to send the corresponding sampling switch trigger parameters to the mission sensor module when the pose compensation item meets the preset action trigger threshold.

6. The unmanned aerial vehicle (UAV) autonomous inspection system based on historical inspection visual features according to claim 1, characterized in that, The dynamic parameter sensing module is used to acquire pitch angle data, roll angle data, and yaw angle data of the flight platform body module at a sampling frequency higher than 50Hz.

7. The unmanned aerial vehicle (UAV) autonomous inspection system based on historical inspection visual features according to claim 3, characterized in that, The multidimensional state estimation submodule is also used to monitor the intensity of changes in ambient irradiance, and when the intensity of changes in ambient irradiance exceeds the preset threshold of 1000 lx, causing abnormalities in the normalized matching residuals, the measurement noise term corresponding to the positioning result in the Kalman filter algorithm is increased.

8. The unmanned aerial vehicle (UAV) autonomous inspection system based on historical inspection visual features according to claim 1, characterized in that, The visual decoupling feature elements in the historical decoupling feature storage module are data elements stored after the three-dimensional rotation inverse transformation is completed on the absolute horizontal reference plane by the field of view coordinate system decoupling transformation module during the preset teaching and inspection phase.

9. The unmanned aerial vehicle (UAV) autonomous inspection system based on historical inspection visual features according to claim 1, characterized in that, The system also includes a redundant safety control module, which is used to drive the power drive control module to switch the flight platform body module to the stationary vertical landing mode when the signal integrity monitoring module determines that the external navigation signal has been continuously lost for more than the preset 300s time limit.

Images