Unmanned aerial vehicle adaptive positioning method and system for mountain photovoltaic power station

Abstract

The present application relates to the technical field of unmanned aerial vehicle positioning and navigation, and particularly relates to an unmanned aerial vehicle adaptive positioning method and system for a mountain photovoltaic power station, the method comprising the following steps: acquiring real-time dynamic differential positioning data, an inversion type of the real-time dynamic differential positioning data, inertial measurement data, downward-looking image data and barometric altitude data; determining a main positioning mode according to the inversion type and a flight phase, the main positioning mode comprising a real-time dynamic differential positioning mode, a terrain matching mode and a visual inertial odometer mode; generating absolute positions or relative positions as observation values in different modes; based on an extended Kalman filter, driving state prediction with the inertial measurement data, and performing state updating with current observation values, adaptively adjusting noise covariance of the observation values according to the flight phase, and outputting continuous and stable unmanned aerial vehicle positioning results. The present application realizes full-range continuous and high-precision positioning of unmanned aerial vehicles in a mountain photovoltaic power station scene, and has strong robustness and engineering practicability.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) positioning and navigation technology, and specifically to an adaptive positioning method and system for UAVs used in mountain photovoltaic power stations. Background Technology

[0002] Drones are widely used in various operations of mountain photovoltaic power plants due to their ability to traverse complex terrains, such as component transportation and inspection. The reliability and safety of drones are highly dependent on continuous and high-precision positioning capabilities. Currently, the mainstream high-precision positioning method is real-time dynamic differential positioning (RTK) technology, which can provide high positioning accuracy in open environments. However, the construction environment of mountain photovoltaic power plants has significant unique characteristics: from the transfer point at the foot of the mountain to the working surface at the top, drones need to pass through areas with steep slopes, dense vegetation, and drastic terrain undulations; at the mountainside, canyon obstruction and multipath effects often cause RTK signal loss or degradation to floating solutions; at the photovoltaic array area at the top of the mountain, the strong reflection of the photovoltaic panel surface will interfere with satellite signals. Therefore, relying solely on RTK cannot guarantee the continuity of positioning throughout the entire flight.

[0003] In existing technologies, visual inertial odometry (VIO) can provide relative positioning in the absence of satellite signals, but it suffers from inherent drift over time, with accumulated errors reaching several meters after long-distance flights. Other visual repositioning methods based on pre-stored maps can provide absolute position correction, but in large mountainous areas, relying solely on natural features results in computationally intensive repositioning and its success rate is significantly affected by ambient lighting. Therefore, there is an urgent need for a UAV positioning method that can adapt to mountainous photovoltaic power station scenarios, achieving adaptive integration of multi-source data to ensure positioning continuity and accuracy throughout the entire flight. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention aims to provide an adaptive positioning method for unmanned aerial vehicles (UAVs) used in mountain photovoltaic power stations, which can achieve continuous and highly robust positioning output throughout the entire flight of the UAV.

[0005] The basic solution provided by this invention is an adaptive positioning method for unmanned aerial vehicles (UAVs) used in mountain photovoltaic power stations, comprising the following steps: S1: Acquire the real-time dynamic differential positioning data and its solution type, inertial measurement data, nadir image data and barometric altitude data of the UAV; and determine the current flight stage of the UAV based on the barometric altitude data and preset waypoints; S2: Based on the solution type and the flight phase, determine the primary positioning mode currently being used from multiple positioning modes; the multiple positioning modes include real-time dynamic differential positioning mode, terrain matching mode, and visual inertial odometry mode; S3: Based on the primary positioning mode, generate observations in the following manner: When the primary positioning mode is real-time dynamic differential positioning mode, the real-time dynamic differential positioning data is used as the observation value; when the primary positioning mode is terrain matching mode, the absolute position of the UAV is calculated by matching the features of the bottom view image with the pre-stored terrain map, and the result is used as the observation value; when the primary positioning mode is visual inertial odometry mode, visual inertial odometry calculation is performed based on the inertial measurement data and the image features of the bottom view image, and the relative position between adjacent frames is output as the observation value. S4: Based on the extended Kalman filter, state prediction is driven by the inertial measurement data, and state updates are performed using the observations obtained from the current primary positioning mode; wherein, the noise covariance of the observations is adaptively adjusted according to the flight phase; and a continuous and stable UAV positioning result is output.

[0006] Furthermore, the flight phase includes the takeoff and climb phase, the mountainside cruise phase, the summit descent phase, and the hovering and dropping phase.

[0007] Furthermore, in S2, the current primary positioning mode is determined based on the solution type and the flight phase, and executed in the following priority order: First priority: When the solution type is a fixed solution, the real-time dynamic differential positioning mode is adopted; Second priority: When the solution type is a floating solution or a single-point solution, and the current flight phase is the hillside cruise phase or the mountaintop descent phase, if the number of interior points obtained after feature matching exceeds a preset threshold, the terrain matching is determined to be successful, and the terrain matching mode is switched; otherwise, the current positioning mode is maintained. Third priority: When the solution type is no solution, if the terrain matching is successful, the terrain matching mode is adopted; if the terrain matching fails, the visual inertial odometry mode is switched; otherwise, the current positioning mode remains unchanged.

[0008] Furthermore, the pre-stored terrain map is generated in the following way: the bottom view image of the construction area is collected in advance by drone and the corresponding real-time dynamic differential positioning data is recorded to construct the terrain map. The terrain map contains the three-dimensional coordinates of natural feature points and their corresponding feature descriptors, and the photovoltaic array layout information is pre-stored.

[0009] Furthermore, the terrain matching mode defaults to the natural feature matching sub-mode; when the drone's flight altitude is below a preset threshold and its current location is within a preset photovoltaic array area, it switches to the photovoltaic array-specific sub-mode. The logic for generating observations in the natural feature matching sub-mode includes: retrieving candidate regions in a pre-stored terrain map with the most recent valid positioning location as the center; extracting feature points and their feature descriptors from the current bottom view image, matching them with the feature descriptors of map feature points in the candidate regions, and using a pose calculation algorithm to calculate the pose of the UAV relative to the map coordinate system to obtain the absolute position as the observation value. The logic for generating observations under the dedicated sub-mode for photovoltaic arrays includes: using a target detection model to identify the corner points of the photovoltaic panel frame in the downward view image, matching the detected photovoltaic panel layout with the photovoltaic array layout information in the pre-stored terrain map, and using the two-dimensional and three-dimensional correspondence of the corner points to calculate the UAV pose and obtain the absolute position as the observation value.

[0010] Furthermore, S4 also includes an inertial drift compensation mechanism: flight data of the construction area is collected in advance, and a recurrent neural network is trained to predict the drift compensation amount relative to the visual inertial odometry calculation result in the future based on the inertial measurement data of the past period; in visual inertial odometry mode, the predicted drift compensation amount is used to correct the prediction state of the extended Kalman filter.

[0011] Furthermore, the state vector of the extended Kalman filter includes position, velocity, attitude quaternions, accelerometer bias, and gyroscope bias; its state prediction step is driven by inertial measurement data, and its state update step receives corresponding observations based on the current master positioning mode.

[0012] The principle of this invention lies in constructing a multi-level adaptive mechanism with the goal of outputting continuous and stable positioning results. First, by dividing the entire operation of the UAV into four flight phases, the signal quality, visual environment, and positioning accuracy requirements of the dynamic differential positioning data are different in each phase. Based on this, a three-layer main positioning mode is designed: dynamic differential positioning mode, terrain matching mode, and visual inertial odometry mode. The mode is switched according to the solution type and flight phase with a fixed priority. This design ensures that at least one mode with acceptable accuracy is available under any signal conditions, thereby ensuring uninterrupted positioning.

[0013] Secondly, since the noise characteristics of different observation sources change with the environment, this invention directly uses the flight phase as the adjustment parameter of the observation noise covariance of the extended Kalman filter. That is, the same observation value is assigned different covariances at different phases, thereby changing the degree of trust of the filter in it. This makes it possible for the weights to be adaptively adjusted even when using the same sensor or algorithm, thus improving the smoothness and accuracy of positioning.

[0014] Furthermore, regarding the specific implementation of the terrain matching mode, this invention sets up a dual-mode system: a default natural feature matching (suitable for textured areas such as hillsides) and a photovoltaic-specific corner point matching system automatically triggered when entering a preset photovoltaic area at low altitude (suitable for weak textures and reflective environments). The two sub-modes share the same offline map and complement each other.

[0015] Finally, to address the long-distance drift problem inherent in visual inertial odometry modes, this invention trains a recurrent neural network offline to predict and compensate for drift online, significantly extending the effective flight distance of the visual inertial odometry mode and maintaining acceptable positioning accuracy even when no absolute position correction is available.

[0016] The beneficial effects of this invention are as follows: First, to ensure continuous positioning, a three-level primary positioning mode switching system is constructed: existing technologies either rely on a single dynamic differential positioning data, which results in large errors or even direct interruption when the signal is lost; or simply switch to visual inertial odometry for positioning after the dynamic differential positioning data fails. This invention, based on the solution type of the dynamic differential positioning data and the flight phase, determines the primary positioning mode according to a clear priority. When the dynamic differential positioning mode degrades, the system prioritizes terrain matching (which can provide high-precision absolute position). Only when terrain matching fails consecutively does the visual inertial odometry relative positioning method activate. This progressive design ensures that the positioning source is always the most reliable under the current conditions, thus maintaining continuous and uninterrupted accurate positioning output throughout the entire flight.

[0017] Secondly, by introducing noise covariance adjustment into the extended Kalman filter during the flight phase, dynamic adaptation of the trust level of the observation source is achieved: Conventional methods set fixed noise parameters for each sensor, which cannot adapt to the drastic fluctuations in signal quality and imaging conditions in mountainous environments. However, this invention introduces noise covariance adjustment into the extended Kalman filter during the flight phase, achieving dynamic adaptation of the trust level of the observation source. According to the four phases of takeoff and climb, mountain cruise, mountain descent, and hovering deployment, different noise covariances are assigned to the observation values ​​output by the current main positioning mode. For example, the covariance of photovoltaic matching is reduced during the mountain descent phase (more trust in vision), while a smaller covariance is assigned to the data during the open cruise phase (more trust in satellite). This design allows the filter to always fuse inertial measurement data (IMU) predictions and observation values ​​with the most reasonable trust level without manual parameter adjustment, resulting in smoother positioning results and higher accuracy.

[0018] Furthermore, for the special scenario of mountain photovoltaic, a two-layer matching mechanism of natural feature matching and photovoltaic-specific corner point matching is designed: conventional visual relocalization relies on natural features, which is prone to failure in the weak texture and reflective environment of photovoltaic arrays. This invention designs a two-layer terrain matching mechanism of natural feature matching and photovoltaic-specific corner point matching for the special scenario of mountain photovoltaic. The coordinates of natural feature points and photovoltaic panel corner points are saved in the offline map. When running online, natural feature matching is used by default. Once the height is detected to be below the threshold and the preset photovoltaic area is entered, it automatically switches to photovoltaic-specific sub-mode. This sub-mode uses the target detection model to identify corner points and uses the artificial regularity of the photovoltaic array to calculate the pose, which is more accurate.

[0019] Finally, an online inertial drift compensation based on a recurrent neural network is introduced for the visual inertial odometry mode. Since long-distance drift is an inherent drawback of this method, conventional approaches require frequent loop closure detection or external reference point correction. This invention trains a lightweight gated recurrent network by pre-collecting inertial measurement data and the actual position during mountain flight. This network learns the mapping relationship between the timing of inertial measurement data and the amount of drift. During online operation in visual inertial odometry mode, this network is activated, directly compensating the predicted drift into the extended Kalman filter (EKF) prediction state. This data-driven method can capture systematic errors that are difficult to model precisely, such as bias residuals and calibration errors in inertial measurement data, enabling the visual inertial odometry mode to maintain a low drift rate even in the absence of absolute positioning correction.

[0020] This invention also discloses an adaptive positioning system for unmanned aerial vehicles (UAVs) used in mountain photovoltaic power stations, comprising: The sensor module is used to acquire real-time dynamic differential positioning data and its solution type, inertial measurement data, downward view image data, and barometric altitude data; The map storage module is used to store a pre-generated topographic map of the construction area. The topographic map includes the three-dimensional coordinates of natural feature points and their corresponding feature descriptors, as well as photovoltaic array layout information. The mode switching module is used to determine the current primary positioning mode based on the solution type and the flight phase, and generate corresponding observations according to the following rules: when the primary positioning mode is real-time dynamic differential positioning mode, the real-time dynamic differential positioning data acquired by the sensor module is used as the observation; when the primary positioning mode is terrain matching mode, the feature matching of the bottom view image acquired by the sensor module and the terrain map in the map storage module is performed to calculate the absolute position as the observation; when the primary positioning mode is visual inertial odometry mode, visual inertial odometry is calculated based on the inertial measurement data acquired by the sensor module and the image features of the bottom view image, and the relative position between adjacent frames is output as the observation. The state estimation module employs an extended Kalman filter to drive state prediction based on the inertial measurement data, update the state using the observations obtained from the current primary positioning mode, and adaptively adjust the noise covariance of the observations according to the flight phase, outputting continuous and stable UAV positioning results. Attached Figure Description

[0021] Figure 1 This is a flowchart of an embodiment of the UAV adaptive positioning method for mountain photovoltaic power stations according to the present invention. Detailed Implementation

[0022] The following detailed description illustrates the specific implementation method: The example uses a drone inspection mission at a mountain photovoltaic power station as an example. The inspection route goes from the transfer point at the foot of the mountain to the working face at the top of the mountain, crossing the photovoltaic array area on the mountainside and the top of the mountain.

[0023] Hardware configuration: The UAV is equipped with: a real-time dynamic differential positioning receiver (outputs positioning data and its solution type, including fixed solution, floating solution, single-point solution or no solution); an inertial measurement unit (outputs three-axis acceleration and three-axis angular velocity); a global shutter downward-looking camera (outputs downward-looking image); a barometric altimeter (outputs barometric altitude data); and an onboard computer for executing the positioning method of this invention.

[0024] Offline preparation: (1) Terrain map construction: Before construction, the same drone is used to cover the entire inspection path and surrounding area along a preset route (such as a "bow" shape), flying at an appropriate altitude to ensure that the forward overlap rate and lateral overlap rate meet the requirements of 3D reconstruction. During the flight, the real-time dynamic differential positioning fixed solution coordinates and attitude of each frame of the bottom view image are recorded to generate a 3D sparse point cloud. The 3D coordinates of natural feature points (such as ORB features) and their corresponding feature descriptors are extracted to construct a terrain map with spatial index. At the same time, for the photovoltaic array area on the mountaintop, the 3D coordinates of the corner points of the photovoltaic panels are manually marked, and the layout information is generated and stored in the terrain map.

[0025] (2) Inertial drift compensation network training: Inertial measurement data (three-axis acceleration and three-axis angular velocity) and true position (obtained by real-time dynamic differential positioning post-processing) from multiple round-trip flights are collected to construct a recurrent neural network. In this embodiment, a gated recurrent unit (GRU) network is selected. The network input is a sequence of inertial measurement data within a continuous time window in the past, and the output is the drift compensation amount (three-dimensional position drift) relative to the visual inertial odometry calculation result in the future. The network is trained using supervised learning so that it can learn the mapping relationship between inertial measurement data and visual inertial odometry drift.

[0026] As attached Figure 1 As shown, the entire online operation process is as follows: S1. Data Acquisition and Flight Phase Identification: After the drone takes off, the onboard computer acquires in real time: real-time dynamic differential positioning data and its solution type, inertial measurement data, drooping image data, and barometric altitude data; Based on barometric altitude data and preset waypoints, the current flight phase of the UAV is divided into: takeoff and climb phase, hillside cruise phase, summit descent phase, and hovering and deployment phase. The takeoff and climb phase is the process of ascending from the takeoff point to the cruise altitude; the hillside cruise phase is the process of horizontally traversing the mountainous area within the preset altitude range; the summit descent phase is the process of descending from the cruise altitude to the hovering and deployment altitude, typically accompanied by entering the photovoltaic array area; the hovering and deployment phase is the process of hovering within a small area above the target point in preparation for deploying the photovoltaic modules. The requirements for positioning accuracy, real-time performance, and robustness vary across different flight phases, and the reliability of each sensor also changes with each phase.

[0027] S2, Mode Switching Decision: Based on the solution type of the real-time dynamic differential positioning data and the current flight phase, the primary positioning mode is determined in the following priority order (three primary positioning modes: real-time dynamic differential positioning mode, terrain matching mode, and visual inertial odometry mode): First priority: When the solution type is a fixed solution, the real-time dynamic differential positioning mode is adopted. At this time, the satellite signal quality is optimal, and high-precision absolute position can be obtained directly, which is the highest precision positioning source. Second priority: When the solution type is a floating solution or a single-point solution, and the current flight phase is the hillside cruising phase or the mountaintop descent phase, attempt terrain matching. Specifically, extract feature points from the current bottom view image and match them with map feature points in the pre-stored terrain map. If the number of interior points obtained after feature matching exceeds a preset threshold, the terrain matching is considered successful, and the system switches to terrain matching mode; otherwise, the current positioning mode is maintained (i.e., no forced switching is performed to avoid frequent jumps due to brief matching failures). Third priority: When the solution type is no solution, first try terrain matching. If terrain matching is successful, then use terrain matching mode; if terrain matching fails (e.g., feature extraction is difficult or map retrieval does not find a match), then switch to visual inertial odometry mode, relying on inertial measurement data and the downward view image to provide relative positioning; For other cases not covered by the above priorities (e.g., the solution type is floating but in the takeoff and climb phase), the current positioning mode is kept unchanged to avoid unnecessary mode jumps; during the mode switching process, the output positioning results are processed by weighted moving average smoothing to eliminate possible jumps at the moment of switching and ensure the stability of flight control.

[0028] S3. Generate observations based on the primary positioning mode: Based on the primary localization mode determined by S2, the observations of the extended Kalman filter are generated as follows: When the primary positioning mode is real-time dynamic differential positioning mode: the real-time dynamic differential positioning data (position, velocity) is directly used as the observation value, and this observation value is the absolute position; When the primary positioning mode is terrain matching mode: feature matching is performed between the bottom view image and the pre-stored terrain map to calculate the absolute position of the UAV as the observation value. The terrain matching mode is further divided into two sub-modes: Natural Feature Matching Sub-mode (default active): Centered on the most recent valid location, candidate regions are retrieved from the pre-stored terrain map. Feature points and their feature descriptors of the current bottom view image are extracted and matched with the feature descriptors of map feature points in the candidate regions. The pose calculation algorithm (such as PnP) is used to calculate the pose of the UAV relative to the map coordinate system, and the absolute position is obtained as the observation value. This sub-mode is suitable for areas with rich natural textures, such as hillsides.

[0029] Photovoltaic array-specific sub-mode (automatically switched when the UAV's flight altitude is below a preset threshold and its current position is within a preset photovoltaic array area): The target detection model (such as YOLOv8-nano) is used to identify the corner points of the photovoltaic panel frame in the downward view image. The detected photovoltaic panel layout is matched with the photovoltaic array layout information in the pre-stored terrain map. The UAV pose is calculated using the two-dimensional and three-dimensional correspondence of the corner points to obtain the absolute position as the observation value. This sub-mode is specifically designed for the weak texture and reflective environment of photovoltaic arrays and has higher robustness.

[0030] When the primary positioning mode is visual inertial odometry mode: visual inertial odometry is calculated based on inertial measurement data and image features of the downward-view image. Specifically, by extracting salient feature points in consecutive image frames and establishing inter-frame correspondences, and combining the pre-integration of inertial measurement data, the relative position (i.e., position change) of the UAV between adjacent frames is calculated, and this relative position is used as the observation value. It should be noted that this observation value is only a relative change, not an absolute coordinate. However, within the extended Kalman filter framework, combined with the absolute position state of the previous moment, the current absolute position estimate can be recursively obtained.

[0031] S4. Adaptive estimation based on extended Kalman filter: The state vector of the extended Kalman filter includes position, velocity, attitude quaternions, accelerometer bias, and gyroscope bias. The filter always operates at a high frequency, and its operation consists of two steps: prediction and update. State prediction: Driven by inertial measurement data, using the acceleration and angular velocity from the inertial measurement data, the predicted state and covariance at the current moment are recursively derived from the state at the previous moment through kinematic equations. This step is performed at a high frequency (e.g., 200Hz) to ensure smooth and continuous filter output; State Update: The state is updated based on the observations obtained from the current primary positioning mode. The update frequency depends on the type of observation source (approximately 20Hz for real-time dynamic differential positioning data, approximately 1-5Hz for successful terrain matching, and approximately 30Hz for a single frame of visual inertial odometry). During each update, the filter uses only the unique observation corresponding to the current primary positioning mode. If the observation value is an absolute position (real-time dynamic differential positioning or terrain matching), it is directly compared with the predicted position for correction; if the observation value is a relative position (visual inertial odometry), it is corrected through the observation model (the difference between the current predicted position and the position of the previous state should be consistent with the observation value).

[0032] Adaptive adjustment of noise covariance: The noise covariance of the current observation is dynamically adjusted according to the current flight phase. The same observation source has different noise characteristics in different flight phases. For example, during the takeoff and climb phase and the open area cruise phase, the real-time dynamic differential positioning signal is stable, so its observation noise covariance is set to be smaller (i.e., given higher confidence), while the covariance of visually related observations is set to be larger. During the descent phase at the summit, the real-time dynamic differential positioning data may jump due to reflection from the photovoltaic panels. At the same time, the drone's altitude decreases and the downward view image becomes clearer. At this time, the observation noise covariance of the terrain matching (photovoltaic-specific sub-mode) is set to a small value (i.e., more trust is placed in visual positioning), and the real-time dynamic differential positioning data is not used or its covariance is set to a large value. During the hovering deployment phase, the absolute position accuracy requirement is the highest. The covariance of the photovoltaic-specific matching is further compressed to ensure high-precision output. Through this adaptive adjustment, the filter can always fuse inertial predictions and observations with the optimal level of confidence without the need for manual parameter tuning.

[0033] Output: The final output is a continuous and stable UAV positioning result (position, velocity, attitude), which is used by the flight control system for navigation and control.

[0034] Inertial drift compensation mechanism: In visual inertial odometry mode, in order to suppress cumulative drift during long-distance flight, this embodiment integrates an inertial drift compensation mechanism. The offline-trained recurrent neural network (such as a GRU network) is invoked when running online: the network predicts the drift compensation amount of the visual inertial odometry calculation result relative to the true value in the future period based on the inertial measurement data in a continuous time window in the past; the predicted drift compensation amount is directly added to the position component of the predicted state of the extended Kalman filter, and the predicted state covariance is increased accordingly to reflect the uncertainty of the compensation. This compensation is only activated in visual inertial odometry mode.

[0035] The UAV adaptive positioning system for mountain photovoltaic power stations, which performs the above method, provided in this embodiment includes: Sensor module: Used to acquire real-time dynamic differential positioning data and its solution type, inertial measurement data, downward view image data, and barometric altitude data.

[0036] Map storage module: Used to store pre-generated terrain maps of the construction area. The map contains the three-dimensional coordinates of natural feature points and their corresponding feature descriptors, as well as photovoltaic array layout information.

[0037] Mode switching module: Used to determine the current primary positioning mode based on the solution type and flight phase, and provide corresponding observations according to the following rules: When the primary positioning mode is real-time dynamic differential positioning mode, the real-time dynamic differential positioning data acquired by the sensor module is used as the observation value. When the main positioning mode is terrain matching mode, the lower view image obtained by the sensor module is used to perform feature matching with the terrain map in the map storage module, and the absolute position is calculated as the observation value. When the main positioning mode is visual inertial odometry mode, visual inertial odometry calculation is performed based on the inertial measurement data acquired by the sensor module and the image features of the bottom view image, and the relative position between adjacent frames is output as the observation value.

[0038] State estimation module: It adopts an extended Kalman filter to drive state prediction with inertial measurement data, updates the state with the observations provided by the mode switching module, and adaptively adjusts the noise covariance of the observations according to the flight phase, outputting continuous and stable UAV positioning results.

[0039] It should be noted that when the primary positioning mode is switched to visual inertial odometry mode, the state vector of the extended Kalman filter already contains the absolute position at the time of the switch (from the successful update of the last real-time dynamic differential positioning or terrain matching). Using this absolute position as a reference, the filter can continuously output absolute position estimates through prediction driven by inertial measurement data and updates of relative position observations.

[0040] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. An adaptive positioning method for unmanned aerial vehicles (UAVs) used in mountain photovoltaic power stations, characterized in that, Includes the following steps: S1: Acquire the real-time dynamic differential positioning data and its solution type, inertial measurement data, nadir image data and barometric altitude data of the UAV; and determine the current flight stage of the UAV based on the barometric altitude data and preset waypoints; S2: Based on the solution type and the flight phase, determine the primary positioning mode currently being used from multiple positioning modes; the multiple positioning modes include real-time dynamic differential positioning mode, terrain matching mode, and visual inertial odometry mode; S3: Based on the primary positioning mode, generate observations in the following manner: When the primary positioning mode is real-time dynamic differential positioning mode, the real-time dynamic differential positioning data is used as the observation value; when the primary positioning mode is terrain matching mode, the absolute position of the UAV is calculated by matching the features of the bottom view image with the pre-stored terrain map, and the result is used as the observation value; when the primary positioning mode is visual inertial odometry mode, visual inertial odometry calculation is performed based on the inertial measurement data and the image features of the bottom view image, and the relative position between adjacent frames is output as the observation value. S4: Based on the extended Kalman filter, state prediction is driven by the inertial measurement data, and state is updated using the observations obtained from the current primary positioning mode; wherein, the noise covariance of the observations is adaptively adjusted according to the flight phase. Output continuous and stable UAV positioning results.

2. The UAV adaptive positioning method for mountain photovoltaic power stations according to claim 1, characterized in that, The flight phases include the takeoff and climb phase, the hillside cruise phase, the summit descent phase, and the hovering and dropping phase.

3. The UAV adaptive positioning method for mountain photovoltaic power stations according to claim 2, characterized in that, S2 describes determining the current primary positioning mode based on the solution type and the flight phase, and executing the process in the following priority order: First priority: When the solution type is a fixed solution, the real-time dynamic differential positioning mode is adopted; Second priority: When the solution type is a floating solution or a single-point solution, and the current flight phase is the hillside cruise phase or the mountaintop descent phase, if the number of interior points obtained after feature matching exceeds a preset threshold, the terrain matching is determined to be successful, and the terrain matching mode is switched; otherwise, the current positioning mode is maintained. Third priority: When the solution type is no solution, if the terrain matching is successful, the terrain matching mode is adopted; if the terrain matching fails, the visual inertial odometry mode is switched; otherwise, the current positioning mode remains unchanged.

4. The UAV adaptive positioning method for mountain photovoltaic power stations according to claim 1, characterized in that, The pre-stored terrain map is generated in the following way: the bottom view image of the construction area is collected in advance by drone and the corresponding real-time dynamic differential positioning data is recorded to construct the terrain map. The terrain map contains the three-dimensional coordinates of natural feature points and their corresponding feature descriptors, and the photovoltaic array layout information is pre-stored.

5. The UAV adaptive positioning method for mountain photovoltaic power stations according to claim 4, characterized in that, The terrain matching mode defaults to the natural feature matching sub-mode; when the drone's flight altitude is below a preset threshold and its current location is within the preset photovoltaic array area, it switches to the photovoltaic array dedicated sub-mode. The logic for generating observations in the natural feature matching sub-mode includes: retrieving candidate regions in a pre-stored terrain map with the most recent valid positioning location as the center; extracting feature points and their feature descriptors from the current bottom view image, matching them with the feature descriptors of map feature points in the candidate regions, and using a pose calculation algorithm to calculate the pose of the UAV relative to the map coordinate system to obtain the absolute position as the observation value. The logic for generating observations under the dedicated sub-mode for photovoltaic arrays includes: using a target detection model to identify the corner points of the photovoltaic panel frame in the downward view image, matching the detected photovoltaic panel layout with the photovoltaic array layout information in the pre-stored terrain map, and using the two-dimensional and three-dimensional correspondence of the corner points to calculate the UAV pose and obtain the absolute position as the observation value.

6. The UAV adaptive positioning method for mountain photovoltaic power stations according to claim 1, characterized in that, S4 also includes an inertial drift compensation mechanism: flight data of the construction area is collected in advance, and a recurrent neural network is trained to predict the drift compensation amount relative to the visual inertial odometry calculation result in the future based on the inertial measurement data of the past period; in visual inertial odometry mode, the predicted drift compensation amount is used to correct the prediction state of the extended Kalman filter.

7. The UAV adaptive positioning method for mountain photovoltaic power stations according to claim 1, characterized in that, The state vector of the extended Kalman filter includes position, velocity, attitude quaternions, accelerometer bias, and gyroscope bias; Its state prediction step is driven by inertial measurement data, and its state update step receives the corresponding observations based on the current primary positioning mode.

8. An adaptive positioning system for unmanned aerial vehicles (UAVs) used in mountain photovoltaic power stations, characterized in that: For performing the method according to any one of claims 1 to 7, comprising: The sensor module is used to acquire real-time dynamic differential positioning data and its solution type, inertial measurement data, downward view image data, and barometric altitude data; The map storage module is used to store a pre-generated topographic map of the construction area. The topographic map includes the three-dimensional coordinates of natural feature points and their corresponding feature descriptors, as well as photovoltaic array layout information. The mode switching module is used to determine the current primary positioning mode based on the solution type and the flight phase, and generate corresponding observations according to the following rules: when the primary positioning mode is real-time dynamic differential positioning mode, the real-time dynamic differential positioning data acquired by the sensor module is used as the observation; when the primary positioning mode is terrain matching mode, the feature matching of the bottom view image acquired by the sensor module and the terrain map in the map storage module is performed to calculate the absolute position as the observation; when the primary positioning mode is visual inertial odometry mode, visual inertial odometry is calculated based on the inertial measurement data acquired by the sensor module and the image features of the bottom view image, and the relative position between adjacent frames is output as the observation. The state estimation module employs an extended Kalman filter to drive state prediction based on the inertial measurement data, update the state using the observations obtained from the current primary positioning mode, and adaptively adjust the noise covariance of the observations according to the flight phase, outputting continuous and stable UAV positioning results.

Images