High-precision topographic surveying and mapping system and method based on unmanned aerial vehicle

Abstract

The invention discloses a high-precision topographic surveying and mapping system and method based on an unmanned aerial vehicle, and particularly relates to the technical field of unmanned aerial vehicle surveying and mapping, and the method comprises the following steps: carrying out the topographic modeling before a task, recognizing GNSS shielding, and dividing a navigation region; gNSS, INS and visual inertial navigation positioning information are fused in flight, the weight is dynamically adjusted, and historical trajectory constraint is combined; constructing a closed-loop detection mechanism to carry out error feedback and feedforward prediction; confidence is calculated according to image definition, IMU stability and time synchronization deviation, high confidence points are screened for DEM reconstruction, local heavy flight is triggered by low confidence points, and surveying and mapping precision and reliability are improved; according to the method, the autonomous navigation and positioning precision of the unmanned aerial vehicle in a complex environment is improved, the quality control of surveying and mapping data and the DEM reconstruction reliability are enhanced, the closed-loop self-correction and feed-forward error prediction capabilities are provided, and the stability, precision and intelligent level of the system are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of UAV mapping, and more specifically, to a high-precision terrain mapping system and method based on UAVs. Background Art

[0002] In existing UAV terrain mapping systems, a combination of GNSS (Global Navigation Satellite System) and inertial navigation system (INS) is often used for flight positioning and attitude solution. However, in complex terrain environments, such as forest-covered areas, canyon areas, or urban high-rise building groups, GNSS signals are often severely blocked, resulting in a decrease in satellite positioning accuracy or an inability to obtain positioning information. At this time, the system usually relies on INS for short-term inertial compensation navigation.

[0003] However, in the absence of external correction signals, the inertial navigation system will have error accumulation in the integration solution process based on acceleration and angular velocity, resulting in continuous attitude and position drift. When GNSS occlusion and INS drift occur simultaneously, the positioning error caused by their superposition will rapidly expand, causing a systematic offset of the geographical coordinates of the mapping results.

[0004] This systematic offset will directly affect the matching relationship between point cloud data, image data, and geographical coordinates, and then lead to obvious deformation, dislocation, or high distortion of the finally generated digital elevation model (DEM), making the obtained mapping results unusable for subsequent engineering design and geographical information analysis, and seriously reducing the reliability and applicability of the mapping system in complex environments. Therefore, the present invention proposes a high-precision terrain mapping system and method based on UAVs to solve the above problems. Summary of the Invention

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A high-precision terrain mapping method based on UAVs, comprising the following steps:

[0007] Step 1: Before the start of the mapping task, obtain remote sensing images or historical terrain data, perform terrain modeling on the target area, identify the GNSS signal occlusion level based on deep learning algorithms, and divide it into GNSS available areas and GNSS restricted areas accordingly;

[0008] Step 2: During the flight in the GNSS available area, fuse GNSS, INS, and visual inertial navigation positioning information, dynamically adjust the weights of each positioning source according to the GNSS signal-to-noise ratio and data integrity, and during the flight in the GNSS restricted area, enable INS and visual inertial navigation collaborative navigation, and use historical high-confidence GNSS trajectory data to perform error constraint on the positioning results;

[0009] Step 3: Based on the flight attitude obtained by sensor fusion, a closed-loop detection mechanism is constructed. When repeatedly passing through the same scene area, image and point cloud comparison is performed, and the detected error is fed back to the inertial navigation system to correct attitude drift. At the same time, the future path error is predicted through a time series model for feedforward compensation;

[0010] Step 4: The image and point cloud data are evaluated and classified according to the confidence index. The confidence value is calculated by weighting based on image clarity, IMU stability, and multi-sensor time synchronization accuracy. Then, a high-confidence subset is selected for digital elevation model reconstruction, and the low-confidence subset is subjected to local re-flight to improve the modeling quality.

[0011] In a preferred embodiment, in Step 1, a convolutional neural network is combined with a terrain semantic segmentation algorithm, and an identification model is trained based on a terrain slope map, a building density map, and a historical GNSS signal-to-noise ratio heat map to perform multi-dimensional semantic classification on the GNSS occlusion risk area, output an occlusion level label, and divide the GNSS available area and the GNSS restricted area.

[0012] In a preferred embodiment, in Step 2, within the GNSS signal available area, the GNSS signal-to-noise ratio, the number of satellites, and the positioning residual value are collected in real time, and a fuzzy logic function is used to calculate the GNSS confidence factor. According to the mapping result of the GNSS confidence factor value within the preset value range, the participation weights of GNSS, INS, and visual inertial navigation in the fusion algorithm are adjusted.

[0013] In a preferred embodiment, in Step 2, the multi-source positioning data is fused and updated in real time through a fusion algorithm, namely a Kalman filter.

[0014] In a preferred embodiment, within the GNSS restricted area, an INS + visual inertial navigation collaborative navigation mechanism is adopted. By simultaneously accessing inertial sensor data and monocular / binocular image streams, optimization registration is performed based on the forward visual odometry of the reference frame and the current frame IMU trajectory, and an "inertial navigation - vision" joint state space model is constructed using the optimized trajectory to achieve high-frequency positioning updates.

[0015] In a preferred embodiment, the reference trajectory in the historical GNSS high-confidence area is composed of data recorded in multiple previous tasks in the same area. A trajectory clustering algorithm is used to extract the center line of the confidence trajectory, and a Kalman filter is used to constrain and correct the current INS trajectory.

[0016] In a preferred embodiment, the closed-loop detection mechanism refers to: running synchronously based on image similarity matching and point cloud geometric reconstruction comparison. For the image part, key frame re-identification is performed based on local descriptors. For the point cloud part, the local overlap rate and geometric centroid offset are calculated as error indicators, and then a comprehensive judgment is made on whether the current closed-loop error exceeds the expectation.

[0017] In a preferred embodiment, in step three, when correcting the attitude of the inertial navigation system, an error feedback adjustment mechanism is adopted. The offset vector obtained from the closed-loop detection of the image and the point cloud is input into the INS resolver in the form of discrete control quantities as fine-tuning quantities to iteratively update the quaternion attitude estimation value.

[0018] The time series model is a time series regression model based on a gated recurrent neural network. The input is the sequence of position error vectors of the previous N frames and the corresponding attitude change rates, and the output is the error prediction trajectory of the next M frames, which is used for future path pre-compensation adjustment.

[0019] In a preferred embodiment, the confidence index includes weighted calculation of three items: image clarity score, IMU dynamic stability value, and sensor time synchronization deviation to obtain a confidence value. The point cloud data with a confidence value greater than or equal to the preset confidence threshold is divided into the high-confidence subset, and the point cloud data with a confidence value less than the preset confidence threshold is divided into the low-confidence subset.

[0020] In a preferred embodiment, a high-precision terrain mapping system based on an unmanned aerial vehicle includes:

[0021] Terrain modeling module: used to obtain remote sensing images or historical terrain data before the mapping task starts, perform terrain modeling on the target area, and identify the GNSS signal occlusion level based on deep learning algorithms, and then divide the area into GNSS available areas and GNSS restricted areas;

[0022] Fusion navigation module: used to fuse GNSS, INS, and visual inertial navigation positioning information during flight in GNSS available areas, and dynamically adjust the weights of each positioning source according to the GNSS signal-to-noise ratio and data integrity. During flight in GNSS restricted areas, INS and visual inertial collaborative navigation are enabled, and the positioning result is error-constrained by combining historical high-confidence GNSS trajectory data;

[0023] Error detection module: used to construct a closed-loop detection mechanism based on the flight attitude obtained by sensor fusion, compare the image and the point cloud when repeatedly passing through the same scene area, and feedback the detected error to the inertial navigation system to correct the attitude drift;

[0024] Error prediction module: used to predict future path errors through a time series model and use the prediction results for feed-forward compensation of the flight path;

[0025] Data evaluation module: used to perform quality evaluation and classification on the collected images and point cloud data according to the confidence index, and the confidence value is calculated by weighted calculation based on image clarity, IMU stability, and multi-sensor time synchronization accuracy;

[0026] Reconstruction processing module: used to screen out high-confidence subsets to participate in the digital elevation model reconstruction, and perform local re-flight operations on low-confidence subsets to improve the overall quality of terrain modeling.

[0027] Technical effects and advantages of the present invention:

[0028] By enabling the inertial navigation system and the visual inertial navigation system to work together in areas with severely blocked GNSS signals, the present invention significantly improves the positioning ability of the unmanned aerial vehicle under the condition of no satellite positioning. In particular, by constructing a "inertial navigation-vision" joint state space model based on the optimized registration of image frames and IMU trajectories, and combining historical GNSS high-confidence trajectory data for error constraint, it is possible to effectively suppress the drift error of the inertial navigation system in complex terrain environments, realize the dynamic maintenance of the navigation accuracy of the unmanned aerial vehicle, and thus ensure that the terrain mapping task can still be carried out stably and continuously under the condition of lack of GNSS signals.

[0029] The present invention introduces a multi-dimensional confidence index system. By weighted fusion of three indicators: image clarity, IMU dynamic stability, and multi-sensor time synchronization deviation, the data confidence of each survey point is calculated in real time, and the point cloud data is classified and managed accordingly. High-confidence data is used for the main model reconstruction, and low-confidence data triggers the local re-flight mechanism, thus completing the quality control at the data source. This mechanism can effectively avoid problems such as model distortion, misalignment, and error diffusion caused by low-quality data, and greatly improve the accuracy and consistency of the finally generated digital elevation model (DEM).

[0030] The present invention constructs a closed-loop detection mechanism based on image similarity and point cloud geometric comparison, which can actively identify the closed-loop error of the flight trajectory when the unmanned aerial vehicle repeatedly passes through the same scene area, and feedback the detected offset vector to the inertial navigation system in the form of discrete control quantity for attitude correction. At the same time, combined with a time series prediction model based on a gated recurrent neural network, feedforward compensation control of future path errors is realized. The above mechanism not only realizes the closed-loop self-correction ability of the system, but also enhances the intelligent prediction ability of path planning, and improves the adaptability, stability and intelligence of the entire terrain mapping system in a dynamic environment. Description of the Drawings

[0031] For the convenience of those skilled in the art to understand, the present invention will be further described below in conjunction with the drawings;

[0032] Figure 1Schematic diagram of the high-precision terrain mapping method based on an unmanned aerial vehicle in the present invention.

[0033] Figure 2 Schematic diagram of the principle of step two in the present invention.

[0034] Figure 3 Schematic diagram of the high-precision terrain mapping system based on an unmanned aerial vehicle in the present invention. Detailed implementation manners

[0035] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0036] Refer to Figure 1 - Figure 3 The following embodiments are obtained:

[0037] Embodiment 1:

[0038] The high-precision terrain mapping method based on an unmanned aerial vehicle includes the following steps:

[0039] Step 1: Before the mapping task starts, obtain remote sensing images or historical terrain data, perform terrain modeling on the target area, identify the GNSS signal occlusion level based on a deep learning algorithm, and divide it into a GNSS available area and a GNSS restricted area accordingly;

[0040] Step 2: During the flight in the GNSS available area, fuse GNSS, INS, and visual inertial navigation positioning information, dynamically adjust the weights of each positioning source according to the GNSS signal-to-noise ratio and data integrity. During the flight in the GNSS restricted area, enable INS and visual inertial collaborative navigation, and perform error constraint on the positioning result by combining historical high-confidence GNSS trajectory data;

[0041] Step 3: Based on the flight attitude obtained by sensor fusion, construct a closed-loop detection mechanism, perform image and point cloud comparison when passing through the same scene area repeatedly, and feed back the detected error to the inertial navigation system to correct the attitude drift. At the same time, predict the future path error through a time series model for feedforward compensation;

[0042] Step 4: Evaluate and classify the image and point cloud data according to the confidence index. The confidence value is calculated by weighting based on image clarity, IMU stability, and multi-sensor time synchronization accuracy. Then, select the high-confidence subset for digital elevation model reconstruction, and perform local re-flight on the low-confidence subset to improve the modeling quality.

[0043] In step 1, a convolutional neural network is combined with a terrain semantic segmentation algorithm to train an identification model based on a terrain slope map, a building density map, and a historical GNSS signal-to-noise ratio heat map, perform multi-dimensional semantic classification on the GNSS occlusion risk area, output occlusion level labels, and divide the GNSS available area and the GNSS restricted area. The specific steps are as follows:

[0044] In this step, to achieve accurate perception and hierarchical management of the GNSS signal occlusion degree in complex terrains, a fusion model of a convolutional neural network (CNN) and a terrain semantic segmentation algorithm is adopted to support intelligent flight path planning and navigation strategy formulation in the mission area. The specific sub-steps are as follows:

[0045] Step 1.1: Obtain prior data of the target survey area, and obtain various basic geographical information of the survey area from public or authorized data sources, including but not limited to:

[0046] Terrain slope map: Reflects the surface slope distribution, and is used to evaluate the terrain undulation and possible signal occlusion situations;

[0047] Building distribution map: Marks the positions, densities, and heights of buildings in the area, and is used to identify GNSS occlusion obstacles in urban scenes;

[0048] Historical GNSS signal-to-noise ratio heat map: Drawn based on the signal-to-noise ratio information recorded by GNSS devices in historical missions, and reflects the GNSS availability of each area. These data are processed by a geographic information system (GIS) and uniformly projected onto a unified coordinate system to ensure spatial consistency for subsequent model input.

[0049] Step 1.2: Input the above data into a multi-modal deep neural network model. The multi-modal deep neural network model is composed of a convolutional neural network (CNN) as the backbone structure and receives multiple prior data modalities as input channels. The model architecture includes the following components: The input channels are respectively connected to the slope map, the building density map, and the signal-to-noise ratio heat map; A multi-scale convolutional structure is used to extract terrain semantic features at different levels; A fusion semantic segmentation module classifies and judges each pixel point in the input area, and outputs a segmentation image representing the occlusion level. This model is trained through a supervised learning method, using historical annotation data as labels, and the training objective is to minimize the occlusion level prediction error.

[0050] Step 1.3: Divide the occlusion level into four levels. The pixel points in the semantic segmentation image output by the model are divided into the following four types of occlusion levels: Completely unoccluded: The GNSS signal is unobstructed and the field of view is unoccluded; Slightly occluded: Partially occluded but the GNSS can be normally received; Moderately occluded: The occlusion degree is relatively high and the GNSS performance degrades; Heavily occluded: In the highly occluded area, the GNSS almost fails. Each level corresponds to a set of navigation fault tolerance parameters, including the maximum allowable flight error, the sensor data fusion frequency, and the available positioning source type, etc., for adapting to the subsequent navigation algorithm.

[0051] Step 1.4: Use the recognition result as a constraint condition and input it into the path planner. The recognition result is passed into the path planner in the form of a mask map or a spatial classification map as the spatial constraint condition for the route generation algorithm. The path planner sets the following parameters in different regions according to the occlusion level: Flight altitude: Increase the flight altitude in the heavily occluded area to avoid occlusion; Flight speed: Reduce the flight speed in the medium and low confidence areas to improve the data sampling quality; Heading angle: Optimize the angle to avoid the high occlusion direction; Sensor activation frequency: Increase the sensor data recording frequency and overlap rate in the occluded area to enhance the later modeling ability. This path planning mechanism enables the UAV to dynamically adapt to the GNSS availability difference and improve the navigation stability and mapping accuracy of the entire route.

[0052] In Step 2, within the available area of the GNSS signal, the SNR, number of satellites, and positioning residual value of the GNSS are collected in real time, and the fuzzy logic function is used to calculate the GNSS confidence factor. According to the mapping result of the GNSS confidence factor value within the preset value range, the participation weights of the GNSS, INS, and visual inertial navigation VIO in the fusion algorithm are adjusted.

[0053] To achieve precise positioning of the UAV within the available area of the GNSS signal and improve the stability and reliability of the navigation solution, a multi-parameter joint evaluation of the GNSS signal quality is adopted, and the participation weights of each sensor in the multi-source positioning fusion algorithm are dynamically adjusted accordingly. The specific process includes the following sub-steps:

[0054] Step 2.1: Collect GNSS signal parameters in real time. The following signal evaluation parameters are obtained through the GNSS in real time: Signal-to-noise ratio (SNR): Reflects the clarity and intensity of the currently received signal; Number of visible satellites (N): The total number of GNSS satellites currently observed; Geometric dilution of precision GDOP: Measures the impact of the satellite geometric distribution on the positioning accuracy, and is obtained by calculating the uniformity (i.e., standard deviation) of the current satellite geometric distribution; Positioning residual RE: The difference value between the actual ranging error and the model calculation value; The above parameters reflect the comprehensive availability of the GNSS in the current environment.

[0055] Step 2.2: Calculate the GNSS confidence factor according to the following fuzzy logic function: Introduce the following fuzzy logic weight function to jointly evaluate the GNSS confidence based on multiple GNSS quality indicators:

[0056] ;

[0057] denotes the GNSS confidence factor, is a stability constant to prevent the denominator from being zero, and its value is set to 0.01; this function ensures that when GDOP or RE increases significantly, automatically decreases to reflect the trend of signal degradation; the output value of this function is always within the range of [0,1], with good normalization characteristics, facilitating the processing of downstream algorithms.

[0058] Step 2.3: According to the mapping result of the value within the range of [0,1], adjust the participation weights of GNSS, INS, and VIO in the fusion algorithm. When the value is higher, the system has a higher dependence on GNSS data. Conversely, the system will automatically reduce the GNSS weight and enhance the dependence on INS and visual inertial navigation (VIO). The weight allocation strategy can be set according to preset rules. For example, if > 0.8, GNSS is dominant, and INS and VIO are used as supplements; if 0.5 < W ≤ 0.8, parallel fusion of GNSS and INS is adopted; if ≤ 0.5, reduce the participation of GNSS and enhance the proportion of INS+VIO fusion navigation. This allocation mechanism can adapt to changes in the GNSS signal environment in real time and enhance the anti-interference and stability capabilities of the navigation system.

[0059] In Step 2, the multi-source positioning data is updated in real time through the fusion algorithm, i.e., the Kalman filter. The aim is to improve the positioning accuracy and stability of the UAV in the area where GNSS signals are available. Through the fusion algorithm, i.e., the Kalman filter, the multi-source positioning data is updated in real time to achieve high-precision and high-confidence navigation solution output.

[0060] During the flight of the drone, the system synchronously receives positioning-related data from multiple sensors. Specifically, it includes the position and velocity information provided by the GNSS module, the integration results of acceleration and angular velocity provided by the inertial navigation system (INS), and the relative displacement and attitude changes calculated by visual-inertial odometry (VIO). The above three types of data are collected synchronously in time series and uniformly converted into the input format of the state parameters of the navigation system. The fusion algorithm uses the Kalman filter as the core data fusion framework. First, the system uses the acceleration and angular velocity data provided by the inertial navigation system, combined with the flight control parameters, to predict and estimate the current state. Subsequently, the system compares the actually collected GNSS and visual-inertial odometry observations with the predicted state, calculates the prediction error, and accordingly corrects the current positioning state.

[0061] In each fusion process, the Kalman filter assigns weights by referring to the observation errors and reliabilities of various sensors. In areas with strong GNSS signals and small positioning errors, the data weight will be increased; while in areas with GNSS occlusion or degraded signal quality, the participation of INS and VIO will be enhanced to achieve dynamic adaptive fusion. After each fusion update, the system will output the fusion solution of the current position, velocity, and attitude, and feedback the correction value to update the internal state variables, forming a closed-loop navigation solution mechanism, so as to achieve continuous, smooth, and high-precision tracking of the drone's motion state. By performing this step, the system can effectively utilize the complementary advantages of multi-source positioning information, significantly reduce the impact caused by the failure of a single sensor or error accumulation, and improve the operation reliability of the entire high-precision terrain mapping system and the accuracy of mapping results.

[0062] In the GNSS-restricted area, an INS + visual-inertial collaborative navigation mechanism is adopted. By simultaneously accessing inertial sensor data and monocular / binocular image streams, based on the forward visual odometry of the reference frame and the current frame IMU trajectory for optimal registration, and using the optimized trajectory to construct an "inertial-visual" joint state space model, high-frequency positioning updates are achieved. In the GNSS signal-restricted environment, through the collaborative processing of the inertial navigation system (INS) and the visual-inertial navigation system (VIO), high-frequency, continuous, and high-precision drone positioning capabilities are realized. This collaborative navigation process includes the following sub-steps:

[0063] Step 3.1: The IMU collects triaxial acceleration and angular velocity data and integrates them to obtain an initial attitude estimate. The inertial navigation system, through the built-in IMU (Inertial Measurement Unit), real-time collects the acceleration data and angular velocity data of the drone in three-dimensional space, and records the linear acceleration and angular velocity information in three directions respectively. Subsequently, the system processes these data through integral operations to obtain a preliminary attitude estimation result of the drone, including position, velocity, and attitude angles (such as roll angle, pitch angle, and yaw angle). This process provides a short-term pose solution with high frequency but error accumulation characteristics.

[0064] Step 3.2: A monocular or binocular camera collects image frames, and the relative displacement between frames is obtained through the Visual-Inertial Odometry (VIO) method. The visual-inertial navigation system collects continuous image frames through a monocular or binocular camera module and processes the image sequence using the Visual-Inertial Odometry (VIO) method. VIO extracts key feature points in the images and tracks their movement trajectories between consecutive frames, thereby calculating the relative displacement and attitude changes that occur to the drone between frames. The results of VIO usually have the characteristics of small drift and long-term stability, but the calculation frequency is limited by the camera frame rate.

[0065] Step 3.3: Construct a joint state space model. The state vector includes position, velocity, attitude, and sensor offsets. The output results of the IMU and VIO are structurally integrated to form a joint state space model. The state vector of this model not only includes the conventional three-dimensional spatial position and velocity vectors but also includes the current attitude angles (or quaternion representation) and the systematic errors (i.e., biases) of the IMU sensors as part of the variables, enabling the entire model to reflect the actual dynamics of the sensor system. By constructing a joint state vector, the VIO observation data and IMU predictions can be processed simultaneously in a unified mathematical space, thus achieving dynamic fusion estimation.

[0066] Step 3.4: Minimize the following cost function to optimize the state estimation. To achieve the optimal fusion between the IMU predictions and VIO observations, the system establishes an optimization objective function , and its form is:

[0067] ;

[0068] represents the actual observation value obtained by the IMU at the i-th moment; represents the IMU value predicted by the system at the i-th moment; represents the actual observation value obtained by the visual odometry at the i-th moment; represents the estimated value of the system for VIO at the i-th moment; Denotes the square of the Euclidean distance, which is used to measure the error magnitude between the predicted value and the observed value; λ is a preset weighting factor for VIO observation residuals, which is used to balance the influence of two types of observations, IMU and VIO, in the optimization; n represents the number of observation frames within the optimization period. By minimizing this objective function, the comprehensive error is minimized, thereby obtaining an optimal trajectory estimation result that conforms to the data trends of both IMU and VIO simultaneously.

[0069] After completing the above optimization, the system will obtain a set of fused state vector results, which is the current joint positioning solution. This result includes the position, velocity, attitude, and sensor bias parameters of the UAV at the current moment, and has higher accuracy and robustness. This joint solution is used as the output of the current navigation system for functions such as path planning, stability control, and map construction of the system. At the same time, the system uses this state as the initial value for the next round of fusion calculation, forming a closed-loop iterative mechanism to ensure continuous, high-frequency, and strong anti-interference positioning during the entire navigation process. The optimized trajectory result is input into the joint state space model. The state vector of this model not only includes position and velocity, but also includes attitude information (such as Euler angles or quaternion representation forms) and IMU sensor offsets, etc. Specifically, the state variables cover the following contents: three-dimensional spatial position; three-dimensional velocity components; three-axis attitude angles or direction cosine matrices; IMU biases (acceleration bias and gyroscope bias); camera-IMU extrinsic parameters (i.e., the relative pose between the two sensors); This joint model unifies visual information and inertial information into the same solution framework, allowing the system to estimate and update the complete state at any moment.

[0070] Since the IMU has a high sampling rate (such as 200Hz or higher), the joint state space model can perform a state prediction every time an IMU data arrives and continuously update the pose estimation during the visual frame interval. When an image frame arrives, a complete optimization update step is performed to correct the trajectory and reduce the error caused by IMU drift. This fusion method enables the system to maintain high-frequency and high-continuity positioning capabilities in an environment without GNSS signals, thereby ensuring that the UAV can still complete precise navigation and terrain mapping tasks in an occluded environment.

[0071] In GNSS-constrained areas, to improve the reliability and accuracy of the inertial navigation trajectory, this system introduces the reference trajectory of the historical GNSS high-confidence area as a constraint condition to correct the current INS trajectory. This method uses the flight trajectory data that has been verified as high-precision in historical tasks to establish a reference path and corrects the error through a fusion algorithm, thereby reducing the influence of inertial navigation error accumulation and improving the navigation solution accuracy. The specific steps are as follows:

[0072] Step 4.1: Extract all historical mission trajectory data in this area from the historical database. The system accesses the historical database containing flight record data, locates the geographical coordinate range corresponding to the current survey area, and retrieves the flight mission trajectory data executed in this area in the past. These trajectory data are from the flight records with strong GNSS signals and small positioning errors in historical missions, and have high spatial accuracy. After verification, they can be used as a stable reference basis.

[0073] Step 4.2: Use the density clustering algorithm (DBSCAN) to screen out the center lines of trajectories in repeatedly overlapping areas.

[0074] After performing unified coordinate transformation and time standardization processing on the extracted historical trajectory data, use the density clustering algorithm, i.e., the density-based spatial clustering algorithm (DBSCAN), to cluster the trajectories. This algorithm can identify the flight path segments that are repeatedly passed through. By the clustering results, outlier trajectories and noise data are removed, and only the areas with dense and highly repetitive trajectories are retained for further extraction of high-confidence trajectories.

[0075] Step 4.3: Fit the clustering results to form a reference trajectory. Perform trajectory fitting on the clustering trajectory results obtained in Step 4.2 to form a continuous and smooth trajectory center line. The fitting process can use methods such as spline curve fitting and least squares fitting to restore the scattered trajectory point set to a reference trajectory line with continuity and stability. This trajectory line represents the path segment with the highest coincidence degree and the smallest navigation error in multiple historical flights in space, and has strong credibility and reference value.

[0076] Step 4.4: Use the Kalman filter method to fuse the current INS estimated trajectory with the reference trajectory and correct the pose error. In the current mission, there are certain degrees of drift errors in the position and pose estimation results output by the inertial navigation system (INS) in real time. To reduce such errors, the system calls the Kalman filter method to perform fusion calculation on the INS estimated trajectory and the reference trajectory generated in Step 4.3. During the fusion process, the Kalman filter uses the position state provided by the reference trajectory as the observation input and the current INS solution state as the prediction input. Through the calculation of state prediction and observation residuals, the dynamic correction of the current pose is realized. The fusion result is the precise positioning solution assisted and constrained by the reference trajectory, effectively suppressing the long-term error accumulation of the inertial system and improving the stability and reliability of the navigation system in GNSS-constrained areas.

[0077] It should be noted that: In GNSS-constrained areas, enable the collaborative navigation of INS and visual inertial navigation, and logically sort out the overall technical process of error constraint by combining historical GNSS high-confidence trajectories, and clarify the input-processing-output boundaries each time:

[0078] The core objectives of Steps 3.1 - 3.4 are: to fuse IMU data and the calculation results of image frames to generate a "preliminary positioning solution". The inputs are IMU data: triaxial acceleration and angular velocity; image stream: monocular or binocular continuous image frames; reference frame selection information: for visual odometry alignment. Processing: Steps 3.1 - 3.4 perform trajectory estimation optimization (IMU trajectory + VIO result → minimize the residual cost function); the output result is the optimized combined trajectory, that is: the positioning solution at the current moment. Output: This result serves as the main positioning reference for real-time flight control (with an update rate of up to dozens of Hz or even higher); however, this trajectory is still a "relative positioning solution", that is, based on the cumulative estimation of IMU and image odometry, there may still be systematic drift (especially after long-term flight).

[0079] The core objectives of Steps 4.1 - 4.4 are: to correct the spatial error of the trajectory solution output by Module 1 and suppress the long-term drift trend. The inputs are the combined trajectory output by Module 1 (i.e., "the current INS solution state"); the "reference trajectory" generated by clustering and fitting historical data; the time alignment mechanism ensures that the two are comparable. Processing: Step 5.4 uses a Kalman filter to take the current INS combined trajectory → as the state prediction input; historical reference trajectory points → as the observation input; perform the calculation of prediction and observation residuals, and adjust the state estimate according to the Kalman gain; obtain the correction value of the current trajectory in the "absolute coordinate space". Output: The "high-confidence trajectory" after fusion, with higher global accuracy.

[0080] The closed-loop detection mechanism refers to: running synchronously based on image similarity matching and point cloud geometric reconstruction comparison. For the image part, key frame re-identification is performed based on local descriptors, and for the point cloud part, the local overlap rate and geometric centroid offset are calculated as error metrics, and then comprehensively judge whether the current closed-loop error exceeds the expectation.

[0081] Step 5.1: Select historical key frame images and the current image, and use descriptors based on ORB or SuperPoint for feature matching. The system first real-time identifies the current image frame during the task and calls the image frame with a similar spatial position to the current image or the previously passed image frame from the historical image database as the historical key frame image. Image feature extraction and matching are performed through local image descriptors. The descriptors can be ORB or SuperPoint, both of which have good anti-rotation and scale change capabilities and are suitable for image appearance change scenarios in the actual surveying and mapping environment. The core of image matching is to extract stable feature points and perform corresponding point matching between the historical frame and the current frame to determine whether they are in the same spatial position or a nearby position area.

[0082] Step 5.2: Extract the point cloud at the corresponding position, perform point cloud matching and overlap rate calculation. When the image matching result reaches a certain confidence threshold, the system will trigger the point cloud data processing module synchronized with the image frame acquisition to extract the point cloud data segment corresponding to the image position. Perform registration operations on these point cloud data to detect the spatial overlap degree between the historical point cloud and the current point cloud. The purpose of point cloud matching is to analyze the similarity in geometric structure between the current frame and the historical frame, and to judge the error source and accumulation degree of the actual flight position through point cloud alignment. Overlap rate calculation can be used to quantify the similarity of the point cloud region and provide a spatial basis for subsequent error calculation.

[0083] Step 5.3: Calculate the image similarity index and the point cloud geometric offset index , and the system numerically evaluates the image matching result and the point cloud matching result respectively:

[0084] The image similarity index represents the matching confidence between the historical image frame and the current image frame;

[0085] The point cloud geometric offset index represents the spatial offset error in the point cloud overlap region. Among them, the calculation method of the point cloud offset index is as follows: Suppose there are N pairs of points that are successfully matched in the overlap region of the historical point cloud and the current point cloud; Pk represents the kth point in the historical point cloud; QK represents the corresponding point in the current point cloud that matches Pk; the average value of the sum of the squares of the Euclidean distances between the two is , representing the geometric offset degree of this overlap region. The larger this value is, the more serious the position offset is, and the more necessary it is to perform error correction.

[0086] Step 5.4: Comprehensively judge whether the current closed-loop error exceeds the expectation. If both the image similarity and the point cloud offset index are within a reasonable range, it is considered that no significant drift has occurred in the current closed-loop area. Otherwise, it is considered that significant drift has occurred in the current closed-loop area and attitude correction is required.

[0087] In step three, when the inertial navigation system attitude is corrected, an error feedback adjustment mechanism is adopted, and the offset vector obtained from the closed-loop detection of the image and the point cloud is input into the INS solver in the form of a discrete control quantity as a fine-tuning quantity to iteratively update the quaternion attitude estimation value.

[0088] During flight, the INS system is driven by the IMU for state integration prediction, and at the same time, the state vector is corrected in real time through the visual inertial optimization algorithm; in addition, when the system detects a closed-loop error between the image and the point cloud, the offset vector is converted into a control quantity and injected into the INS solver state update module to achieve micro-scale attitude correction; finally, the navigation solution refers to the historical high-confidence GNSS trajectory when necessary, and the trajectory-level error compensation is performed through the Kalman filter to obtain a stable and continuous mapping and positioning result.

[0089] In an environment where GNSS signals are unavailable or blocked, the Inertial Navigation System (INS) assumes the main positioning responsibility. Based on the three-axis acceleration and angular velocity collected by the IMU, it calculates the three-dimensional position, velocity, and attitude of the UAV through integration. However, the INS has an inherent defect. Without external correction signals, its output will continuously accumulate errors due to sensor errors and integration drift, ultimately leading to the deviation of attitude estimation from the true value. This system introduces an error feedback adjustment mechanism inside the INS solver. This mechanism takes the spatial offset vector extracted from the closed-loop detection results of images and point clouds as the core, and periodically corrects the INS attitude through external observation data to suppress drift.

[0090] Since the system periodically performs image similarity matching and point cloud reconstruction comparison during flight, if it is determined that the current area forms a closed loop with the historical path, error detection is triggered. The result output by this detection is the spatial offset vector, including the position deviation and the attitude drift amount. Among them, the attitude drift can be indirectly obtained by matching the spatial rotation relationship of the point clouds.

[0091] Among them, attitude drift refers to the angular deviation generated during the long-term integration process of the inertial navigation system, manifested as a small rotation error between the acquisition directions of the camera or lidar and the true flight direction. To obtain this attitude drift amount, the system first confirms the spatial closed-loop relationship between the current frame image and the historical key frame through image matching, and then extracts the point cloud data collected at the corresponding moments of the two frames of images for local three-dimensional point cloud registration. The essence of the registration operation is to perform rigid body transformation alignment on two segments of point clouds in three-dimensional space, which includes two components: position translation and attitude rotation.

[0092] By analyzing the registration transformation relationship between the two segments of point clouds in the overlapping area, the system can separate the rotation amount part. This rotation amount represents the relative rotation of the historical point cloud and the current point cloud around a certain axis in the spatial coordinate system, which is the attitude offset of the aircraft caused by inertial navigation errors between two moments. Since the point cloud data contains spatial geometric structures, this rotation amount has high stability and anti-drift characteristics and is suitable for use as the calculation basis for the attitude correction amount.

[0093] The system does not directly calculate the rotation angle, but reflects this rotation relationship through the spatial transformation matrix of the point cloud or its derived forms (such as rotation matrix or attitude quaternion), thereby indirectly obtaining the attitude drift. Finally, the attitude offset result is converted into a correction amount recognizable by the inertial navigation system and injected into the INS solver in a discrete control form to achieve attitude update and drift suppression.

[0094] Convert the attitude drift part into an equivalent attitude adjustment amount, expressed as a discrete control amount. In the case of representing the attitude by quaternions, this control amount can be regarded as a small quaternion increment, which is used to correct the attitude variables recorded in the current INS state. Take this discrete control amount as a correction factor and input it into the INS solver for state correction. During the filter update process, superimpose this control amount to shift the predicted state towards the true trajectory direction, completing the "state fine-tuning driven by external observations". In the filter structure, this operation is equivalent to artificially constructing an observation residual, enabling the system to perform a round of fusion based on this "pseudo-observation value".

[0095] If the current attitude is represented by the quaternion qt and t represents the time, then the new attitude after fine-tuning is: new attitude = current attitude quaternion * error feedback quaternion. That is, without breaking the state recurrence structure, accurately correct the original integration result to avoid problems such as image distortion, mapping distortion, and trajectory drift caused by long-term attitude drift.

[0096] The "discrete control amount form" refers to: converting continuous space error information into an attitude correction input amount that is periodically applied during the attitude solution process of the inertial navigation system, with a fixed time step and update frequency. Specifically, this discrete control amount form has the following technical characteristics: simple structural form: this control amount usually exists in the form of an attitude correction vector (such as a small Euler angle increment or a quaternion increment), and can be inserted into the INS state update process with low computational overhead; controllable time rhythm: the correction operation is discretely triggered at fixed time points of the navigation state update (such as every several frames or every several seconds), rather than continuous input, to avoid over-regulation or oscillation of the system; strong fusion in the correction method: this control amount, as an "external observation fitting correction", is added to the state transition equation or attitude integration process of the INS in the form of a correction term, belonging to a discrete feedback mechanism driven by observation errors and based on non-model perturbations; good interface adaptability: this control amount can be directly used to correct the quaternion variable used for attitude representation in the INS, and iterative attitude update is achieved through quaternion multiplication or small rotation superposition, without affecting the original prediction structure of the inertial navigation system; for example, in the attitude representation based on quaternions, the system calculates the attitude offset generated by the closed-loop detection as a small quaternion, which is injected into the current state quaternion as a discrete control amount for multiplication combination to achieve local rotation correction of the attitude. Since this correction is performed in a discrete form, the system can flexibly control its frequency and amplitude, achieving both error correction and avoiding excessive interference.

[0097] The time series model is a time series regression model based on a gated recurrent neural network. The input is a sequence of the position error vectors of the previous N frames and the corresponding attitude change rates, and the output is the error prediction trajectory of the next M frames, which is used for pre-compensation adjustment of the future path. The input of this time series model is: a sequence of the position error vectors of the previous N frames; the data of the attitude change rates of the corresponding N frames; these input data reflect the evolution trend of the navigation system error and the dynamic response characteristics during the recent flight process. By learning the temporal dependence relationship between the input data, this model extracts the error change pattern and outputs an error prediction trajectory with a length of M, that is, the error estimation results of the next M frames. The output results are used as the predicted error information in the future path and are input into the path control or navigation correction module for: pre-adjusting the flight attitude or the position solution weight; dynamically modifying the path planning parameters; starting a feed-forward control algorithm for error compensation. The feed-forward control algorithm is a prior art and will not be elaborated here. By introducing this neural network model, the system not only has the current error correction ability but also has the ability to predict and respond to potential future errors, greatly enhancing the navigation robustness in the GNSS-constrained environment.

[0098] The confidence index is calculated by weighted calculation of three items: the image clarity score, the IMU dynamic stability value, and the sensor time synchronization deviation, to obtain the confidence value. The point cloud data with the confidence value greater than or equal to the preset confidence threshold is classified into the high-confidence subset, and the point cloud data with the confidence value less than the preset confidence threshold is classified into the low-confidence subset.

[0099] The image clarity score is used to reflect the visual quality of the current frame image, mainly considering whether the image is clear, whether there is motion blur, too dark or overexposed illumination, etc. The acquisition method is as follows: the system performs edge intensity analysis on the image frame and extracts the image gradient value; specifically, the Sobel operator is used to process the image to calculate the amplitude of the pixel edge change in the image; then the edge intensity of the entire image is statistically analyzed, for example, calculating the variance of the edge intensity; the higher the clarity, the stronger the edge change and the higher the score; conversely, the clarity score of a blurred image is lower. This score is normalized to fluctuate between 0 and 1, where 1 indicates that the image is very clear and 0 indicates that the image is extremely unclear.

[0100] The IMU dynamic stability value is used to measure the stability of the inertial sensor data in the current flight state of the UAV, mainly reflecting whether there are severe jitters, accelerations or irregular movements during flight, which usually affect the IMU integration accuracy. The acquisition method is as follows: The system reads the continuous values of the three-axis angular velocity and acceleration in the IMU in real time; takes one second as the time window and calculates the variance of the angular velocity change during this time period; then takes the reciprocal of the sum obtained by adding a very small number such as 0.0001 to this variance, and performs normalization processing; when the angular velocity change is stable, the variance is small and the reciprocal is large, indicating good dynamic stability; when the flight is severe, the variance rises, the stability decreases, and the corresponding score value decreases. Finally, the IMU dynamic stability value is also normalized to between 0 and 1, where 1 represents high stability and 0 represents very unstable.

[0101] The sensor time synchronization deviation is used to evaluate whether the timestamps of multiple sensors such as image sensors, IMUs, and GNSSs are aligned, mainly considering problems such as sampling asynchronization and data delay. The acquisition method is as follows: When the system synchronizes and processes each frame of image and the corresponding IMU data, it records the timestamps of various sensors; calculates the time interval (difference) between the current image frame and the corresponding IMU data; if this difference exceeds the set threshold (such as 20 milliseconds), it is considered that there is a synchronization offset, otherwise it is considered that there is no synchronization offset; uses this difference as the basic index and performs normalization processing to obtain the time synchronization deviation value; the smaller the synchronization deviation, the higher the score; the larger the synchronization deviation, the lower the score. The final result takes the form of "1 minus the synchronization deviation value" and is converted into a "synchronization confidence value", with a normalization range of 0 to 1. If there is no synchronization offset, the "synchronization confidence value" is directly set to 1.

[0102] The above three indicators are weighted and summarized according to the preset weighting ratio to form the final comprehensive confidence value: The weight of the image clarity score is the largest (such as 40%), because image blurring will directly affect point cloud mapping; the weight of the IMU dynamic stability value is moderate (such as 35%), which determines the quality of the point cloud trajectory; the weight of the sensor time synchronization deviation is slightly lower (such as 25%), but it is mainly used for fusion consistency judgment.

[0103] The confidence value formed after weighting is limited to the range of 0 to 1. Set a confidence threshold (such as 0.7) as the baseline for dividing the quality of point cloud data: When the confidence value corresponding to the point cloud data is greater than or equal to this preset confidence threshold, it is included in the high-confidence subset and will participate in the main path mapping and high-precision DEM generation; High-precision DEM refers to: a digital elevation model with high spatial resolution and high vertical accuracy. When the confidence value corresponding to the point cloud data is less than this confidence threshold, it is included in the low-confidence subset, and the subsequent local re-flight correction mechanism is executed.

[0104] Embodiment 2: A high-precision terrain mapping system based on a UAV, including:

[0105] Terrain Modeling Module: It is used to obtain remote sensing images or historical terrain data before the start of a surveying and mapping task, perform terrain modeling on the target area, identify the GNSS signal occlusion level based on deep learning algorithms, and then divide the area into GNSS available areas and GNSS restricted areas;

[0106] Integrated Navigation Module: It is used to integrate GNSS, INS, and visual inertial navigation positioning information during flight in GNSS available areas, and dynamically adjust the weights of each positioning source according to the GNSS signal-to-noise ratio and data integrity. During flight in GNSS restricted areas, INS and visual inertial collaborative navigation are enabled, and the positioning results are error-constrained by combining historical high-confidence GNSS trajectory data;

[0107] Error Detection Module: It is used to construct a closed-loop detection mechanism based on the flight attitude obtained by sensor fusion, compare images and point clouds when repeatedly passing through the same scene area, and feedback the detected errors to the inertial navigation system to correct attitude drift;

[0108] Error Prediction Module: It is used to predict future path errors through a time series model, and use the prediction results for feedforward compensation of the flight path;

[0109] Data Evaluation Module: It is used to evaluate and classify the collected image and point cloud data according to confidence metrics, and the confidence value is calculated by weighting based on image clarity, IMU stability, and multi-sensor time synchronization accuracy;

[0110] Reconstruction Processing Module: It is used to select high-confidence subsets to participate in digital elevation model reconstruction, and perform local re-flight operations on low-confidence subsets to improve the overall quality of terrain modeling.

[0111] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by software simulation of a large amount of collected data to get a formula closest to the real situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0112] In the multiple calculation formulas involved in the present invention, the interference intensity index, impact breadth index, and the parameters such as the number of sub-packets, packet sending time interval, and reception time extension value calculated therefrom are all subjected to unified numerical standardization processing to ensure the dimensional consistency and applicability in the formula calculation process. In order to avoid the problem of dimensional disorder or mismatch in the formula calculation, the present invention uses a normalization method and a standard data preprocessing process to unify the units and reduce the dimensions of the original collected data during the design phase. Specifically, all operating parameters collected by the system, such as signal strength reception value, signal-to-noise ratio, packet loss rate, retransmission rate, etc., are processed by a standardized algorithm based on dimensionless theory and converted into a unified relative value or standard value, thereby eliminating the unit difference and dimensional interference between different parameters, and ensuring that the numerical operations of each variable and adjustment coefficient in the formula have mathematical rationality and physical consistency.

[0113] The parameters such as adjustment coefficient, amplification factor, dominant coefficient, balance coefficient, etc. used in the formula are all based on the data model that has completed unit elimination and dimensionality reduction, and are obtained through a large number of experimental verifications and simulation tests on standardized data sets. The setting of these coefficients not only conforms to the basic principle of dimensional balance in engineering calculations, but is also repeatedly verified through system simulation and laboratory environment to ensure that the reliability and adaptability of the calculation results will not be affected by the inconsistency of data dimensions. Through the above method, the structure of all formulas in the present invention maintains mathematical logic self-consistency and physical dimension consistency, which can ensure that the calculation model has wide applicability and engineering feasibility under different application environments and system configurations.

[0114] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0115] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0116] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0117] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims described above.

Claims

1. A high-precision terrain mapping method based on UAV, characterized in that: The following steps are involved: Step 1: Before the surveying and mapping task begins, remote sensing images or historical terrain data are obtained to perform terrain modeling on the target area. The GNSS signal obstruction level is identified based on the deep learning algorithm, and the area is divided into GNSS available areas and GNSS restricted areas accordingly. Step 2: During the flight in the GNSS available area, the GNSS, INS and visual inertial navigation positioning information are integrated, and the weight of each positioning source is dynamically adjusted according to the GNSS signal-to-noise ratio and data integrity. During the flight in the GNSS restricted area, the INS and visual inertial navigation collaborative navigation are enabled, and the positioning results are constrained by combining the historical high-confidence GNSS trajectory data; Step 3: Build a closed-loop detection mechanism based on the flight attitude obtained by sensor fusion, compare the image and point cloud when repeatedly passing through the same scene area, and feed back the detected error to the inertial navigation system to correct the attitude drift. At the same time, the future path error is predicted through the time series model for feedforward compensation; Step 4: Evaluate and classify the image and point cloud data according to the confidence index. The confidence value is calculated based on the weighted image clarity, IMU stability and multi-sensor time synchronization accuracy. Then, the high-confidence subset is selected for digital elevation model reconstruction, and the low-confidence subset is locally re-flighted to improve the modeling quality.

2. The high-precision terrain mapping method based on an unmanned aerial vehicle according to claim 1 is characterized in that: In step one, a convolutional neural network is combined with a terrain semantic segmentation algorithm. The recognition model is trained based on the terrain slope map, building density map and historical GNSS signal-to-noise ratio heat map to perform multi-dimensional semantic classification of the GNSS obstruction risk area, output the obstruction level label, and divide the GNSS available area and GNSS restricted area.

3. The high-precision terrain mapping method based on an unmanned aerial vehicle according to claim 2 is characterized in that: In step 2, within the GNSS signal available area, the GNSS signal-to-noise ratio, number of satellites, and positioning residual values ​​are collected in real time, and the GNSS confidence factor is calculated using a fuzzy logic function. According to the mapping result of the GNSS confidence factor value within the preset value range, the participation weights of GNSS, INS, and visual inertial navigation in the fusion algorithm are adjusted.

4. The high-precision terrain mapping method based on an unmanned aerial vehicle according to claim 3 is characterized in that: In step 2, the multi-source positioning data is fused and updated in real time through the fusion algorithm, namely the Kalman filter.

5. The high-precision terrain mapping method based on an unmanned aerial vehicle according to claim 4 is characterized in that: In GNSS restricted areas, the INS+visual inertial navigation collaborative navigation mechanism is adopted. By simultaneously accessing the inertial sensor data and monocular / binocular image streams, the forward visual odometer of the reference frame and the IMU trajectory of the current frame are optimized and aligned, and the "INS-vision" joint state space model is constructed using the optimized trajectory to achieve high-frequency positioning updates.

6. The high-precision terrain mapping method based on an unmanned aerial vehicle according to claim 5 is characterized in that: The reference trajectory of the historical GNSS high confidence area is composed of data recorded by multiple missions in the same area in the past. The trajectory clustering algorithm is used to extract the centerline of the confidence trajectory, and the Kalman filter is used to constrain and correct the current INS trajectory.

7. The high-precision terrain mapping method based on an unmanned aerial vehicle according to claim 6, characterized in that: The closed-loop detection mechanism refers to: the image similarity matching and point cloud geometric reconstruction comparison are run simultaneously. The image part uses key frame re-identification based on local descriptors, and the point cloud part calculates the local overlap rate and geometric center of gravity offset as error indicators, and then comprehensively judges whether the current closed-loop error exceeds expectations.

8. The high-precision terrain mapping method based on an unmanned aerial vehicle according to claim 7, characterized in that: In step 3, the inertial navigation system attitude is corrected using an error feedback adjustment mechanism. The offset vector obtained by closed-loop detection of the image and point cloud is input into the INS solver in the form of a discrete control quantity, and is used as a fine-tuning quantity to iteratively update the quaternion attitude estimate. The time series model is a time series regression model based on a gated recurrent neural network. Its input is the position error vector sequence of the previous N frames and the corresponding attitude change rate. Its output is the error prediction trajectory of the next M frames, which is used for future path pre-compensation adjustment.

9. The high-precision terrain mapping method based on an unmanned aerial vehicle according to claim 8, characterized in that: The confidence index includes three weighted calculations: image clarity score, IMU dynamic stability value and sensor time synchronization deviation, to obtain the confidence value. Point cloud data with a confidence value greater than or equal to the preset confidence threshold is divided into a high confidence subset, and point cloud data with a confidence value less than the preset confidence threshold is divided into a low confidence subset.

10. A high-precision terrain mapping system based on an unmanned aerial vehicle, used to implement the high-precision terrain mapping method based on an unmanned aerial vehicle as claimed in any one of claims 1 to 9, characterized in that: include: Terrain modeling module: used to obtain remote sensing images or historical terrain data before the surveying and mapping mission begins, perform terrain modeling on the target area, and identify the GNSS signal obstruction level based on the deep learning algorithm, and then divide the area into GNSS available area and GNSS restricted area; Fusion navigation module: used to fuse GNSS, INS and visual inertial navigation positioning information during flight in GNSS available areas, and dynamically adjust the weight of each positioning source according to the GNSS signal-to-noise ratio and data integrity. During flight in GNSS restricted areas, INS and visual inertial navigation collaborative navigation are enabled, and historical high-confidence GNSS trajectory data are combined to constrain the error of the positioning results; Error detection module: used to build a closed-loop detection mechanism based on the flight attitude obtained by sensor fusion, compare images with point clouds when repeatedly passing through the same scene area, and feed back the detected errors to the inertial navigation system to correct attitude drift; Error prediction module: used to predict future path errors through time series models and use the prediction results for feedforward compensation of the flight path; Data evaluation module: used to evaluate and classify the collected images and point cloud data according to the confidence index. The confidence value is calculated based on the weighted image clarity, IMU stability and multi-sensor time synchronization accuracy. Reconstruction processing module: used to screen out high-confidence subsets to participate in the reconstruction of the digital elevation model, and perform local re-flight operations on low-confidence subsets to improve the overall quality of terrain modeling.

Images