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

Abstract

The present invention discloses a high-precision terrain mapping system and method based on an unmanned aerial vehicle (UAV), which specifically relates to the technical field of UAV mapping. The system and method comprise the following steps: performing terrain modeling and identifying GNSS obstructions before a mission, and dividing the navigation area; integrating GNSS, INS, and visual inertial navigation positioning information during flight, dynamically adjusting weights, and combining historical trajectory constraints; constructing a closed-loop detection mechanism for error feedback and feedforward prediction; calculating confidence based on image clarity, IMU stability, and time synchronization deviation, screening high-confidence points for DEM reconstruction, and triggering local re-flight for low-confidence points, thereby improving mapping accuracy and reliability. The system improves the autonomous navigation and positioning accuracy of UAVs in complex environments, enhances the quality control of mapping data and the reliability of DEM reconstruction, and possesses closed-loop self-correction and feedforward error prediction capabilities, significantly improving the stability, accuracy, and intelligence level of the system.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle (UAV) surveying and mapping technology, and more particularly, to a high-precision terrain surveying and mapping system and method based on an UAV. Background Art

[0002] Existing drone terrain mapping systems often use a combination of GNSS (Global Navigation Satellite System) and Inertial Navigation System (INS) for flight positioning and attitude calculation. However, in complex terrain environments, such as forested areas, canyons, or urban areas with tall buildings, GNSS signals are often severely blocked, resulting in reduced satellite positioning accuracy or even the inability to obtain positioning information. In these situations, the system typically relies on the INS for short-term inertial compensation navigation.

[0003] However, in the absence of external correction signals, the inertial navigation system's integration of acceleration and angular velocity can lead to accumulated errors, resulting in persistent attitude and position drift. When GNSS obstruction and INS drift occur simultaneously, the combined positioning errors can rapidly increase, causing a systematic shift in the geographic coordinates of the surveyed results.

[0004] This systematic offset directly affects the matching relationship between point cloud data, image data, and geographic coordinates, leading to significant deformation, misalignment, or high distortion in the resulting digital elevation model (DEM). This renders the resulting surveying and mapping results unusable for subsequent engineering design and geographic information analysis, severely reducing the reliability and applicability of surveying and mapping systems in complex environments. Therefore, the present invention proposes a high-precision terrain mapping system and method based on drones to address this issue. Summary of the Invention

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

[0006] The high-precision terrain mapping method based on UAV includes the following steps:

[0007] Step 1: Before the surveying and mapping mission begins, remote sensing images or historical terrain data are acquired to perform terrain modeling of the target area. Based on the deep learning algorithm, the GNSS signal obstruction level is identified and the area is divided into GNSS-available and GNSS-restricted areas.

[0008] Step 2: During flight in GNSS-available areas, integrate GNSS, INS, and visual inertial navigation positioning information, and dynamically adjust the weight of each positioning source based on the GNSS signal-to-noise ratio and data integrity. During flight in GNSS-restricted areas, enable collaborative navigation between INS and visual inertial navigation, and combine historical high-confidence GNSS trajectory data to constrain the positioning results.

[0009] Step 3: Build 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. Feedback the detected errors to the inertial navigation system to correct attitude drift. Simultaneously, a time series model is used to predict future path errors for feedforward compensation.

[0010] Step 4: The image and point cloud data are quality-assessed and classified based on confidence indicators. The confidence value is calculated based on a weighted combination of image clarity, IMU stability, and multi-sensor time synchronization accuracy. A high-confidence subset is then selected for digital elevation model reconstruction, and a local re-flight is performed on the low-confidence subset to improve modeling quality.

[0011] In a preferred embodiment, in step one, a convolutional neural network is combined with a terrain semantic segmentation algorithm to train a recognition model 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 into a GNSS restricted area.

[0012] In a preferred embodiment, in step 2, the GNSS signal-to-noise ratio, number of satellites, and positioning residual values are collected in real time within the GNSS signal available area, and the GNSS confidence factor is calculated using a fuzzy logic function. Based on the mapping result of the GNSS confidence factor value within a 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, multi-source positioning data are fused and updated in real time through a fusion algorithm, namely a Kalman filter.

[0014] In a preferred embodiment, in GNSS restricted areas, an INS+visual inertial navigation collaborative navigation mechanism is adopted. By simultaneously accessing inertial sensor data and monocular / binocular image streams, the forward visual odometry of the reference frame and the IMU trajectory of the current frame are optimized and aligned. The optimized trajectory is used to construct an "INS-vision" joint state space model to achieve high-frequency positioning updates.

[0015] In a preferred embodiment, the reference trajectory of the historical GNSS high-confidence area is composed of data recorded from multiple missions in the same area in the past. The centerline of the confidence trajectory is extracted using a trajectory clustering algorithm, and the current INS trajectory is constrained and corrected using a Kalman filter.

[0016] In a preferred embodiment, the closed-loop detection mechanism refers to: based on the simultaneous operation of image similarity matching and point cloud geometric reconstruction comparison, 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.

[0017] In a preferred embodiment, in step 3, the inertial navigation system attitude is corrected using an error feedback adjustment mechanism, and 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;

[0018] 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.

[0019] In a preferred embodiment, the confidence index includes a weighted calculation of three items: image clarity score, IMU dynamic stability value and sensor time synchronization deviation to obtain a confidence value. Point cloud data with a confidence value greater than or equal to a 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.

[0020] In a preferred embodiment, a high-precision terrain mapping system based on a drone includes:

[0021] Terrain Modeling Module: This module is 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 a deep learning algorithm, thereby dividing the area into GNSS-available and GNSS-restricted areas.

[0022] Fusion Navigation Module: This module is used to fuse GNSS, INS, and VINS positioning information during flight in GNSS-available areas, and dynamically adjust the weight of each positioning source based on the GNSS signal-to-noise ratio and data integrity. During flight in GNSS-restricted areas, INS and VINS collaborative navigation are enabled, and historical high-confidence GNSS trajectory data is combined to constrain the positioning results.

[0023] Error Detection Module: This module is used to build a closed-loop detection mechanism based on the flight attitude obtained by sensor fusion. It compares images and point clouds when repeatedly passing through the same scene area and feeds the detected errors back to the inertial navigation system to correct attitude drift.

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

[0025] Data evaluation module: used to evaluate and classify the quality of collected images and point cloud data based on confidence indicators. The confidence values are calculated based on a weighted combination of image clarity, IMU stability, and multi-sensor time synchronization accuracy.

[0026] Reconstruction processing module: used to screen out 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.

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

[0028] This invention significantly improves the positioning capability of drones in the absence of satellite positioning by enabling the coordinated operation of an inertial navigation system (INS) and a visual inertial navigation system (VIS) in areas with severe GNSS signal obstruction. In particular, by constructing a joint INS-VIS state-space model based on optimized registration of image frames and IMU trajectories, and combining historical GNSS high-confidence trajectory data for error constraint, it effectively suppresses the drift error of the INS in complex terrain environments, dynamically maintaining the navigation accuracy of the drone, and thus ensuring that terrain mapping tasks can be carried out stably and continuously even in the absence of GNSS signals.

[0029] This invention introduces a multidimensional confidence index system. By weightedly integrating three metrics—image clarity, IMU dynamic stability, and multi-sensor time synchronization deviation—it calculates the data confidence of each survey point in real time and uses this confidence to classify and manage point cloud data. High-confidence data is used to reconstruct the main model, while low-confidence data triggers a local re-flight mechanism, thus ensuring quality control at the source of the data. This mechanism effectively avoids model distortion, misalignment, and error diffusion caused by low-quality data, significantly improving the accuracy and consistency of the resulting digital elevation model (DEM).

[0030] The present invention constructs a closed-loop detection mechanism based on image similarity and point cloud geometry comparison. It can actively identify the closed-loop error of the flight trajectory when the drone repeatedly passes through the same scene area, and feed back the detected offset vector to the inertial navigation system in the form of discrete control quantities 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 achieved. The above mechanism not only realizes the closed-loop self-correction capability of the system, but also enhances the intelligent prediction capability of path planning, improving the adaptability, stability and intelligence of the entire terrain mapping system in dynamic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0032] Figure 1This is a schematic diagram of the principle of the high-precision terrain mapping method based on drones in the present invention.

[0033] Figure 2 This is a schematic diagram of the principle of step 2 in the present invention.

[0034] Figure 3 This is a schematic diagram of the high-precision terrain mapping system based on drones in the present invention. DETAILED DESCRIPTION

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0036] Reference Figure 1 - Figure 3 The following examples were obtained:

[0037] Example 1:

[0038] The high-precision terrain mapping method based on UAV includes the following steps:

[0039] Step 1: Before the surveying and mapping mission begins, remote sensing images or historical terrain data are acquired to perform terrain modeling of the target area. Based on the deep learning algorithm, the GNSS signal obstruction level is identified and the area is divided into GNSS-available and GNSS-restricted areas.

[0040] Step 2: During flight in GNSS-available areas, integrate GNSS, INS, and visual inertial navigation positioning information, and dynamically adjust the weight of each positioning source based on the GNSS signal-to-noise ratio and data integrity. During flight in GNSS-restricted areas, enable collaborative navigation between INS and visual inertial navigation, and combine historical high-confidence GNSS trajectory data to constrain the positioning results.

[0041] Step 3: Build 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. Feedback the detected errors to the inertial navigation system to correct attitude drift. Simultaneously, a time series model is used to predict future path errors for feedforward compensation.

[0042] Step 4: The image and point cloud data are quality-assessed and classified based on confidence indicators. The confidence value is calculated based on a weighted combination of image clarity, IMU stability, and multi-sensor time synchronization accuracy. A high-confidence subset is then selected for digital elevation model reconstruction, and a local re-flight is performed on the low-confidence subset to improve modeling quality.

[0043] In step 1, a convolutional neural network is combined with a terrain semantic segmentation algorithm to train a recognition model based on terrain slope maps, building density maps, and historical GNSS signal-to-noise ratio heat maps. This model performs multi-dimensional semantic classification of GNSS obstruction risk areas, outputs obstruction level labels, and demarcates GNSS-available areas and GNSS-restricted areas. The specific steps include:

[0044] In this step, to accurately perceive and manage the degree of GNSS signal obstruction in complex terrain, a fusion model of convolutional neural networks (CNN) and terrain semantic segmentation algorithms is used to support intelligent route planning and navigation strategy formulation in the mission area. The specific steps include the following:

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

[0046] Terrain slope map: reflects the distribution of surface slope, used to assess terrain undulations and possible signal obstruction;

[0047] Building distribution map: Marks the location, density, and height of buildings in the area to identify GNSS obstructions in urban scenes;

[0048] Historical GNSS Signal-to-Noise Ratio Heatmap: This map, based on signal-to-noise ratio data recorded by GNSS equipment during historical missions, reflects GNSS availability in various regions. This data is processed using a geographic information system (GIS) and projected to a unified coordinate system, ensuring spatial consistency for subsequent model input.

[0049] Step 1.2: Input the above data into a multimodal deep neural network model. This model, composed of a convolutional neural network (CNN) as its backbone, accepts multiple prior data modalities as input channels. The model architecture includes the following components: Input channels are connected to the slope map, building density map, and signal-to-noise ratio heat map; a multi-scale convolutional structure is used to extract terrain semantic features at different levels; and a semantic segmentation module is integrated to classify each pixel in the input area and output a segmented image representing the occlusion level. The model is trained using supervised learning, using historically annotated data as labels, with the training objective of minimizing the occlusion level prediction error.

[0050] Step 1.3: Classify the obstruction level into four levels. Pixels in the semantic segmentation image output by the model are classified into the following four levels: Completely unobstructed (GNSS signal is clear and the field of view is unobstructed); Mildly obstructed (partially obstructed but GNSS reception is normal); Moderately obstructed (high obstruction with degraded GNSS performance); and Severely obstructed (highly obstructed areas with near-ineffective GNSS). Each level corresponds to a set of navigation tolerance parameters, including maximum allowable flight error, sensor data fusion frequency, and available positioning source types, which are used to adapt subsequent navigation algorithms.

[0051] Step 1.4: Use the recognition results as constraints and input them into the path planner. The recognition results are passed to the path planner in the form of a mask or spatial classification map, serving as spatial constraints for the route generation algorithm. The path planner sets the following parameters in different areas based on the occlusion level: Flight altitude: Increase the flight altitude in heavily occluded areas to avoid occlusion; Flight speed: Reduce the flight speed in low- to medium-confidence areas to improve data sampling quality; Heading angle: Optimize the angle to avoid directions with high occlusion; Sensor activation frequency: Increase the frequency and overlap of sensor data recording in occluded areas to enhance post-processing modeling capabilities. This path planning mechanism enables the drone to dynamically adapt to differences in GNSS availability, improving navigation stability and mapping accuracy throughout the route.

[0052] 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. The GNSS confidence factor is calculated using a fuzzy logic function. Based on the mapping result of the GNSS confidence factor value within the preset value range, the participation weights of GNSS, INS, and visual inertial navigation (VIO) in the fusion algorithm are adjusted.

[0053] To achieve accurate positioning of drones within the GNSS signal availability area and improve the stability and reliability of navigation solutions, a multi-parameter joint evaluation of GNSS signal quality is used to dynamically adjust the weight of each sensor in the multi-source positioning fusion algorithm. 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 in real time through GNSS: Signal-to-noise ratio (SNR): reflects the clarity and strength of the current received signal; Number of visible satellites (N): the total number of GNSS satellites currently observed; Geometric Dilution of Precision (GDOP): measures the impact of satellite geometric distribution on positioning accuracy, obtained by calculating the uniformity of the current satellite geometric distribution, i.e., the standard deviation; Positioning Residual (RE): the difference between the actual ranging error and the model-calculated value; the above parameters reflect the comprehensive availability of GNSS in the current environment.

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

[0056] ;

[0057] represents the GNSS confidence factor, To prevent the stability constant from being zero, the value is set to 0.01; this function ensures that when GDOP or RE increases significantly, It automatically decreases to reflect the trend of signal degradation; the output value of this function is always in the range of [0,1], with good normalization characteristics, which is convenient for downstream algorithm processing.

[0058] Step 2.3: According to The mapping result of the value in the interval [0,1] is used to adjust the participation weights of GNSS, INS, and VIO in the fusion algorithm. The higher the value, the more the system relies on GNSS data. Conversely, the system will automatically reduce the GNSS weight and increase the reliance on INS and visual inertial navigation (VIO). The weight distribution strategy can be set according to the preset rules. For example, if >0.8, GNSS is dominant, INS and VIO are auxiliary; if 0.5 <W ≤0.8, adopt GNSS and INS parallel fusion; if ≤0.5, reducing GNSS participation and increasing 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 using a fusion algorithm, the Kalman filter. This aims to improve the positioning accuracy and stability of the drone within the GNSS signal availability area. This fusion algorithm, the Kalman filter, updates the multi-source positioning data in real time, achieving a high-precision, high-confidence navigation solution output.

[0060] During drone flight, the system synchronously receives positioning-related data from multiple sensors. This includes position and velocity information provided by the GNSS module, acceleration and angular velocity integration results from the inertial navigation system (INS), and relative displacement and attitude changes calculated by the visual inertial navigation system (VIO). These three types of data are collected synchronously in time series and uniformly converted into the navigation system's state parameter input format. The fusion algorithm uses a 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 flight control parameters, to predict and estimate the current state. The system then compares the actual GNSS and VIO observations with the predicted state, calculates the prediction error, and uses this error to correct the current positioning state.

[0061] During each fusion process, the Kalman filter assigns weights based on the observation errors and credibility of various sensors. For areas with strong GNSS signals and small positioning errors, the data weight will be increased; in areas with GNSS obstruction or signal quality degradation, 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, thereby achieving continuous, smooth, and high-precision tracking of the UAV's motion state. Through the execution of this step, the system can effectively utilize the complementary advantages of multi-source positioning information, significantly reduce the impact of single sensor failure or error accumulation, and improve the operational reliability and mapping accuracy of the entire high-precision terrain mapping system.

[0062] In GNSS-restricted areas, an INS+visual inertial navigation collaborative navigation mechanism is used. By simultaneously accessing inertial sensor data and monocular / binocular image streams, an optimized registration is performed based on the forward visual odometry of the reference frame and the IMU trajectory of the current frame. The optimized trajectory is used to construct an "INS-vision" joint state space model to achieve high-frequency positioning updates. In GNSS signal-restricted environments, the collaborative processing of the inertial navigation system (INS) and the visual inertial navigation system (VIO) achieves high-frequency, continuous, and high-precision UAV positioning capabilities. The collaborative navigation process includes the following sub-steps:

[0063] Step 3.1: The IMU collects and integrates triaxial acceleration and angular velocity data to obtain an initial attitude estimate. The inertial navigation system uses the built-in IMU (Inertial Measurement Unit) to collect real-time acceleration and angular velocity data of the drone in three-dimensional space, recording linear acceleration and angular velocity information in three directions. The system then processes this data through integration to obtain a preliminary attitude estimate of the drone, including position, velocity, and attitude angles (such as roll, pitch, and yaw). This process provides a high-frequency, short-term attitude solution that is subject to error accumulation.

[0064] Step 3.2: Image frames are acquired from monocular or binocular images, and relative displacement between frames is obtained using visual odometry (VIO). The visual inertial navigation system acquires continuous image frames using a monocular or binocular camera module and processes the image sequence using visual odometry (VIO). VIO extracts key feature points from the image and tracks their motion between consecutive frames, thereby inferring the relative displacement and attitude changes of the drone between frames. VIO results typically exhibit low drift and long-term stability, but the computation 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 offset. The outputs of the IMU and VIO are structured and integrated to form a joint state-space model. This model's state vector includes not only the conventional three-dimensional position and velocity vectors, but also the current attitude angle (or quaternion representation) and the systematic error (i.e., bias) of the IMU sensor as part of the variables, allowing the entire model to reflect the actual dynamics of the sensor system. By constructing a joint state vector, VIO observations and IMU predictions can be processed simultaneously in a unified mathematical space, 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 prediction value and the VIO observation value, the system establishes an optimization objective function , which has the form:

[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 time i; represents the actual observation value obtained by the visual odometry at the i-th moment; represents the system's estimated value of VIO at the i-th moment; represents the square of the Euclidean distance, which is used to measure the error between the predicted value and the observed value; λ is the preset VIO observation residual weighting factor, which is used to balance the influence of IMU and VIO observations in the optimization; n represents the number of observation frames in the optimization cycle. By minimizing this objective function, the comprehensive error is minimized, thereby obtaining the optimal trajectory estimation result that conforms to the trends of both IMU and VIO data.

[0069] After completing the optimization, the system generates a fused state vector, the current joint localization solution. This solution includes the drone's position, velocity, attitude, and sensor bias parameters at the current moment, offering enhanced accuracy and robustness. This joint solution serves as the navigation system's output and is used to control the system's path planning, stability control, and map building functions. The system also uses this state as the initial value for the next round of fusion calculations, forming a closed-loop iterative mechanism to ensure continuous, high-frequency positioning with strong interference resistance throughout the navigation process. The optimized trajectory results are input into a joint state-space model, whose state vector includes not only position and velocity, but also attitude information (e.g., Euler angles or quaternions) and IMU sensor offsets. Specifically, the state variables include: 3D spatial position; 3D velocity components; 3-axis attitude angles or direction cosine matrices; IMU biases (acceleration and gyroscope biases); and camera-IMU extrinsics (i.e., the relative pose between the two sensors). This joint model unifies visual and inertial information within a single solution framework, allowing the system to estimate and update the complete state at any moment.

[0070] Because the IMU has a high sampling rate (e.g., 200Hz or higher), the joint state-space model can perform a state prediction upon each incoming IMU data point and continuously update the pose estimate during the visual frame interval. When an image frame arrives, a full optimization update step is performed to correct the trajectory and mitigate errors caused by IMU drift. This fusion approach enables the system to maintain high-frequency and continuous positioning capabilities even in environments without GNSS signals, ensuring that the drone can still complete precise navigation and terrain mapping tasks in obscured environments.

[0071] To improve the reliability and accuracy of inertial navigation trajectories in GNSS-restricted areas, this system introduces reference trajectories from historical GNSS high-confidence zones as constraints to correct the current INS trajectory. This method uses high-precision flight trajectory data from historical missions to establish a reference path and uses a fusion algorithm to perform error correction, thereby reducing the impact of accumulated inertial navigation errors and improving navigation solution accuracy. Specifically, it includes the following steps:

[0072] Step 4.1: Extract all past mission trajectory data for the area from the historical database. The system accesses the historical database containing flight log data, locates the geographic coordinate range corresponding to the current survey area, and retrieves trajectory data for previously executed missions within that area. This trajectory data is derived from historical missions with strong GNSS signals and low positioning errors. It has high spatial accuracy and has been verified to serve as a stable reference.

[0073] Step 4.2: Use the density clustering algorithm (DBSCAN) to filter out the center lines of the trajectory of multiple repeated areas.

[0074] After uniform coordinate transformation and time normalization of the extracted historical trajectory data, the trajectories are clustered using a density-based spatial clustering algorithm (DBSCAN). This algorithm identifies recurring flight path segments and uses the clustering results to remove outlier trajectories and noise data, retaining only regions with dense and repetitive trajectories for further extraction of high-confidence trajectories.

[0075] Step 4.3: Fit the clustering results to form a reference trajectory. The clustered trajectory results obtained in Step 4.2 are then fitted to form a continuous, smooth trajectory centerline. This fitting process can employ methods such as spline curve fitting and least squares fitting to restore the scattered trajectory points to a continuous and stable reference trajectory. This trajectory spatially represents the path segment with the highest degree of overlap and the lowest navigation error across multiple historical flights, providing 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 to correct for pose errors. During the current mission, the position and attitude estimates output by the inertial navigation system (INS) in real time exhibit a certain degree of drift error. To reduce this error, the system invokes the Kalman filter method to fuse the INS estimated trajectory with 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. By calculating the residual between the state prediction and the observation, it dynamically corrects the current pose. The fusion result is a precise positioning solution, constrained by the reference trajectory. This effectively suppresses long-term error accumulation in the inertial system and improves the stability and reliability of the navigation system in GNSS-restricted areas.

[0077] It should be noted that in GNSS-restricted areas, the overall technical process of enabling INS and visual inertial navigation collaborative navigation and combining historical GNSS high-confidence trajectories for error constraint is logically sorted out to clarify the input-processing-output boundaries of each time:

[0078] The core goal of steps 3.1-3.4 is to fuse IMU data and image frame calculations to generate a "preliminary positioning solution." The input is IMU data: triaxial acceleration and angular velocity; image stream: continuous monocular or binocular image frames; and reference frame selection information: used for visual odometry alignment. Processing: Steps 3.1-3.4 perform trajectory estimation optimization (IMU trajectory + VIO results → minimize the residual cost function). The output is the optimized joint trajectory, i.e., the positioning solution at the current moment. Output: This result serves as the primary positioning reference for real-time flight control (up to tens of Hz or even higher update rates). However, this trajectory is still a "relative positioning solution," meaning it is based on cumulative estimates from the IMU and image odometry, and may still exhibit systematic drift (especially after long flights).

[0079] The core objective of steps 4.1 through 4.4 is to perform spatial error correction on the trajectory solution output by module 1 to suppress long-term drift. The input is the joint trajectory output by module 1 (i.e., the "current INS solution state") and the "reference trajectory" generated by cluster fitting of historical data. A time alignment mechanism ensures comparability between the two. Processing: Step 5.4 utilizes a Kalman filter to take the current INS joint trajectory as the state prediction input and the historical reference trajectory points as the observation input. The Kalman filter calculates the residual between the prediction and observation, adjusts the state estimate based on the Kalman gain, and obtains the corrected value of the current trajectory in the "absolute coordinate space." The output is a fused "high-confidence trajectory" with higher global accuracy.

[0080] The closed-loop detection mechanism refers to: based on the simultaneous operation of image similarity matching and point cloud geometric reconstruction and comparison, 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.

[0081] Step 5.1: Select the historical keyframe image and the current image, and use ORB- or SuperPoint-based descriptors for feature matching. The system first identifies the current image frame in real time during the task, and calls image frames with similar or previously passed spatial positions to the image from the historical image database as historical keyframe images. Image feature extraction and matching are performed using local image descriptors. The descriptors can be ORB or SuperPoint, both of which have good resistance to rotation and scale changes and are suitable for scenes with image appearance changes in actual surveying and mapping environments. The core of image matching is to extract stable feature points and match corresponding points in the historical frame and the current frame to determine whether they are in the same spatial position or a similar position area.

[0082] Step 5.2: Extract the corresponding point cloud, perform point cloud matching, and calculate the overlap rate. When the image matching result reaches a certain confidence threshold, the system triggers the point cloud data processing module, which is collected synchronously with the image frame, to extract the point cloud data fragment corresponding to the image location. This point cloud data is then registered to detect the degree of spatial overlap between the historical point cloud and the current point cloud. The purpose of point cloud matching is to analyze the geometric similarity between the current frame and the historical frame. Point cloud alignment is used to determine the source and accumulation of errors in the actual flight position. Overlap rate calculation can be used to quantify the similarity of point cloud regions, providing a spatial basis for subsequent error calculations.

[0083] Step 5.3: Calculate image similarity metrics Offset index from point cloud geometry , the system performs numerical evaluation on the image matching results and point cloud matching results 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 overlapping area of the point cloud. The calculation method of the point cloud offset index is as follows: there are N point pairs that are successfully matched in the overlapping area 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 , which indicates the degree of geometric offset of the overlapping area. The larger the value, the more serious the position offset is and the more error correction is needed.

[0086] Step 5.4: Comprehensively judge whether the current closed-loop error exceeds expectations. If the image similarity and point cloud offset indicators are both within a reasonable range, it is considered that the current closed-loop area has not drifted significantly. Otherwise, it is considered that the current closed-loop area has drifted significantly and posture correction is required.

[0087] In step three, 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.

[0088] During flight, the INS system is driven by the IMU to perform state integral prediction, while the state vector is iteratively corrected in real time through a visual inertial navigation 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 variable and injected into the INS solver state update module to achieve micro-scale attitude correction. The final navigation solution refers to historical high-confidence GNSS trajectories when necessary, and trajectory-level error compensation is performed through a Kalman filter to obtain stable and continuous surveying and positioning results.

[0089] In environments where GNSS signals are unavailable or obstructed, the Inertial Navigation System (INS) assumes primary positioning responsibility. Based on the three-axis acceleration and angular velocity collected by the IMU, it infers the drone's three-dimensional position, velocity, and attitude through integration. However, INS has inherent flaws. In the absence of external correction signals, its output will continue to accumulate errors due to sensor errors and integration drift, ultimately causing the attitude estimate to deviate from the true value. This system introduces an error feedback adjustment mechanism within the INS solver. This mechanism, centered on the spatial offset vector extracted from image and point cloud closed-loop detection results, periodically corrects the INS attitude using external observation data to suppress drift.

[0090] During flight, the system periodically performs image similarity matching and point cloud reconstruction comparisons. If a closed loop occurs between the current area and the historical path, an error detection is triggered. This detection output is a spatial offset vector, consisting of position deviation and attitude drift. The attitude drift can be indirectly determined by matching the spatial rotation of the point cloud.

[0091] Among them, attitude drift refers to the angular deviation generated by the inertial navigation system during the long integration process, which manifests as a small rotational error between the camera or lidar acquisition direction and the actual flight direction. To obtain this attitude drift, the system first confirms the existence of a spatial closed-loop relationship between the current frame image and the historical keyframe through image matching. It then extracts the point cloud data collected at the corresponding moments of the two frames and performs local 3D point cloud registration. The essence of the registration operation is to align two point clouds in 3D space by rigid body transformation, which includes two components: position translation and attitude rotation.

[0092] By analyzing the registration transformation relationship between the two point clouds within the overlapping region, the system can isolate the rotational component. This rotational component represents the relative rotation between the historical and current point clouds about a specific axis in the spatial coordinate system, representing the attitude offset of the aircraft caused by inertial guidance error between the two moments. Because the point cloud data contains spatial geometry, this rotational component exhibits high stability and drift resistance, making it suitable for calculating attitude corrections.

[0093] The system does not directly calculate the rotation angle. Instead, it indirectly calculates the attitude drift by reflecting the rotation relationship through the spatial transformation matrix of the point cloud or its derived form (such as the rotation matrix or attitude quaternion). Ultimately, this attitude offset result is converted into a correction value that can be recognized by the inertial navigation system and injected into the INS solver in the form of discrete control to achieve attitude update and drift suppression.

[0094] The attitude drift is converted into an equivalent attitude adjustment, expressed as a discrete control variable. When the attitude is represented by quaternions, this control variable can be considered a tiny quaternion increment, used to correct the attitude variables recorded in the current INS state. This discrete control variable is input into the INS solver as a correction factor for state correction. This control variable is then added to the filter update process, shifting the predicted state toward the true trajectory, completing "external observation-driven state fine-tuning." In the filter structure, this operation is equivalent to artificially constructing an observation residual, allowing the system to perform a round of fusion based on this "pseudo-observation."

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

[0096] The “discrete control quantity form” refers to converting the continuous spatial error information into an attitude correction input with a fixed time step and update frequency that is periodically applied during the attitude solution process of the inertial navigation system. Specifically, this discrete control quantity has the following technical features: A concise structure: This control quantity typically takes the form of an attitude correction vector (such as a micro-increment of Euler angles or a quaternion increment) and can be inserted into the INS state update process with low computational overhead. A controllable timing: The correction operation is discretely triggered at fixed time points during the navigation state update (such as every few frames or seconds), rather than being a continuous input, to avoid system over-adjustment or oscillation. A highly adaptable correction method: This control quantity acts as an "external observation fitting correction" and is added as a correction term to the INS state transition equation or attitude integration process. This is a discrete feedback mechanism driven by observation errors and based on non-model perturbations. A well-adapted interface: This control quantity can be directly used to correct the quaternion variables used for attitude representation in the INS, achieving iterative attitude updates through quaternion multiplication or small rotation superposition without affecting the original prediction structure of the inertial navigation system. For example, in a quaternion-based attitude representation, the system calculates the attitude offset generated by closed-loop detection as a small quaternion, which is then injected into the current state quaternion as a discrete control quantity for multiplication and combination, achieving local rotation correction of the attitude. Since the correction is performed in discrete form, the system can flexibly control its frequency and amplitude, which not only achieves error correction but also avoids excessive interference.

[0097] The time series model is a time series regression model based on a gated recurrent neural network. Its input is the sequence of position error vectors and the corresponding attitude change rates for the previous N frames. Its output is an error prediction trajectory for the next M frames, which is used for pre-compensation adjustment of the future path. The input to this time series model is the sequence of position error vectors for the previous N frames and the corresponding N frames of attitude change rate data. These inputs reflect the navigation system's error evolution trends and dynamic response characteristics during recent flight. The model learns the temporal dependencies between the input data, extracts error variation patterns, and outputs an error prediction trajectory of length M, which is the error estimate for the next M frames. This output serves as the predicted error information for the future path and is input into the path control or navigation correction module. It is used to pre-adjust the flight attitude or position solution weights; dynamically modify path planning parameters; and initiate a feedforward control algorithm for error compensation. The feedforward control algorithm is currently available and will not be further described here. By introducing this neural network model, the system not only has the ability to correct current errors but also predict and respond to potential future errors, significantly enhancing navigation robustness in GNSS-constrained environments.

[0098] The confidence index includes three weighted calculations: image clarity score, IMU dynamic stability value, and sensor time synchronization deviation. The confidence value is obtained. 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.

[0099] The image clarity score reflects the visual quality of the current frame, primarily considering image clarity, motion blur, dark lighting, or overexposure. This score is obtained as follows: the system performs edge strength analysis on the image frame and extracts the image gradient. Specifically, the Sobel operator is used to process the image, calculating the magnitude of pixel edge variation. Edge strength statistics for the entire image are then calculated, for example, by calculating the variance of edge strength. Higher clarity scores are associated with stronger edge variation, while blurred images have lower clarity scores. This score is normalized to fluctuate between 0 and 1, with 1 indicating a very clear image and 0 indicating an extremely unclear image.

[0100] The IMU dynamic stability value is used to measure the stability of the inertial sensor data in the drone's current flight state. It mainly reflects whether there is severe jitter, acceleration, or irregular movement during flight, which usually affects the IMU's 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; using a one-second time window, it calculates the variance of the angular velocity change within this time period; then, after adding an extremely small number such as 0.0001 to the variance, the sum is taken and normalized. When the angular velocity changes smoothly, the variance is small and the reciprocal is large, indicating good dynamic stability. When the flight is intense, the variance increases, the stability decreases, and the corresponding score value decreases. Finally, the IMU dynamic stability value is also normalized to between 0 and 1, with 1 indicating high stability and 0 indicating very unstable.

[0101] Sensor time synchronization deviation is used to assess the alignment of timestamps between multiple sensors, such as image sensors, IMUs, and GNSS sensors. It primarily considers issues such as sampling asynchrony and data latency. This is achieved by recording the timestamps of each sensor during synchronization of each image frame with the corresponding IMU data. The time interval (difference) between the current image frame and the corresponding IMU data is calculated. If this difference exceeds a set threshold (e.g., 20 milliseconds), synchronization deviation is considered present; otherwise, no synchronization deviation is considered present. This difference is then normalized to yield a time synchronization deviation value. Smaller synchronization deviations are associated with higher scores, while larger deviations are associated with lower scores. The final result, calculated as "1 minus the synchronization deviation value," is converted into a "synchronization confidence value," normalized between 0 and 1. If no synchronization deviation exists, the "synchronization confidence value" is assigned a value of 1.

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

[0103] The confidence value formed after weighting is limited to the range of 0 to 1. A confidence threshold (such as 0.7) is set as the baseline for point cloud data quality classification: when the corresponding confidence value of the point cloud data is greater than or equal to the preset confidence threshold, it is classified as a high-confidence subset and will participate in 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 corresponding confidence value of the point cloud data is less than the confidence threshold, it is classified as a low-confidence subset and the local re-fly correction mechanism is subsequently implemented.

[0104] Example 2: A high-precision terrain mapping system based on a drone, comprising:

[0105] Terrain Modeling Module: This module is 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 a deep learning algorithm, thereby dividing the area into GNSS-available and GNSS-restricted areas.

[0106] Fusion Navigation Module: This module is used to fuse GNSS, INS, and VINS positioning information during flight in GNSS-available areas, and dynamically adjust the weight of each positioning source based on the GNSS signal-to-noise ratio and data integrity. During flight in GNSS-restricted areas, INS and VINS collaborative navigation are enabled, and historical high-confidence GNSS trajectory data is combined to constrain the positioning results.

[0107] Error Detection Module: This module is used to build a closed-loop detection mechanism based on the flight attitude obtained by sensor fusion. It compares images and point clouds when repeatedly passing through the same scene area and feeds the detected errors back to the inertial navigation system to correct attitude drift.

[0108] Error prediction module: 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: used to evaluate and classify the quality of collected images and point cloud data based on confidence indicators. The confidence values are calculated based on a weighted combination of image clarity, IMU stability, and multi-sensor time synchronization accuracy.

[0110] Reconstruction processing module: used to screen out 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 numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[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, data packet sending time interval, and reception time extension value calculated therefrom are all subjected to unified numerical standardization processing to ensure their 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, data 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 the adjustment coefficient, amplification factor, dominant coefficient, balance coefficient, etc. used in the formula are all based on a 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 environments to ensure that the reliability and adaptability of the calculation results are not affected by the inconsistency of data dimensions. Through the above method, the structures of all formulas in the present invention maintain mathematical logic self-consistency and consistency in physical dimensions, 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 skilled 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 beyond the scope of this application.

[0116] Those skilled in the art will 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] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or modifications that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A high-precision terrain mapping method based on drones, characterized in that: The following steps are involved: Step 1: Before the surveying and mapping mission begins, remote sensing images or historical terrain data are acquired to perform terrain modeling of the target area. Based on the deep learning algorithm, the GNSS signal obstruction level is identified and the area is divided into GNSS-available and GNSS-restricted areas. Step 2: During flight in GNSS-available areas, integrate GNSS, INS, and visual inertial navigation positioning information, and dynamically adjust the weight of each positioning source based on the GNSS signal-to-noise ratio and data integrity. During flight in GNSS-restricted areas, enable collaborative navigation between INS and visual inertial navigation, and combine historical high-confidence GNSS trajectory data to constrain the positioning results. Step 3: Build 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. Feedback the detected errors to the inertial navigation system to correct attitude drift. Simultaneously, a time series model is used to predict future path errors for feedforward compensation. Step 4: The image and point cloud data are quality-assessed and classified based on confidence metrics. Confidence values are calculated based on a weighted combination of image clarity, IMU stability, and multi-sensor time synchronization accuracy. A high-confidence subset is then selected for digital elevation model reconstruction, while a local re-flight is performed on the low-confidence subset to improve modeling quality. In GNSS-restricted areas, an INS+visual inertial navigation collaborative navigation mechanism is adopted. By simultaneously accessing inertial sensor data and monocular / binocular image streams, an optimized registration is performed based on the forward visual odometry of the reference frame and the IMU trajectory of the current frame. The optimized trajectory is used to construct an "INS-vision" joint state space model to achieve high-frequency positioning updates. The reference trajectory of the historical GNSS high-confidence area is composed of data recorded by multiple missions in the same area. The centerline of the confidence trajectory is extracted using a trajectory clustering algorithm, and the current INS trajectory is constrained and corrected using a Kalman filter. The closed-loop detection mechanism refers to: based on the simultaneous operation of image similarity matching and point cloud geometric reconstruction and comparison, 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.

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

3. The high-precision terrain mapping method based on an unmanned aerial vehicle according to claim 2, 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. The GNSS confidence factor is calculated using a fuzzy logic function. Based on 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, 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, 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 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.

6. The high-precision terrain mapping method based on an unmanned aerial vehicle according to claim 5, characterized in that: The confidence index includes three weighted calculations: image clarity score, IMU dynamic stability value, and sensor time synchronization deviation. The confidence value is obtained. 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.

7. 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 according to any one of claims 1 to 6, characterized in that: include: Terrain Modeling Module: This module is 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 a deep learning algorithm, thereby dividing the area into GNSS-available and GNSS-restricted areas. Fusion Navigation Module: This module is used to fuse GNSS, INS, and VINS positioning information during flight in GNSS-available areas, and dynamically adjust the weight of each positioning source based on the GNSS signal-to-noise ratio and data integrity. During flight in GNSS-restricted areas, INS and VINS collaborative navigation are enabled, and historical high-confidence GNSS trajectory data is combined to constrain the positioning results. Error Detection Module: This module is used to build a closed-loop detection mechanism based on the flight attitude obtained by sensor fusion. It compares images and point clouds when repeatedly passing through the same scene area and feeds the detected errors back to the inertial navigation system to correct attitude drift. Error prediction module: used to predict future path errors through a time series model and use the prediction results for feedforward compensation of the flight path; Data evaluation module: used to evaluate and classify the quality of collected images and point cloud data based on confidence indicators. The confidence values are calculated based on a weighted combination of image clarity, IMU stability, and multi-sensor time synchronization accuracy. Reconstruction processing module: 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; In GNSS-restricted areas, an INS+visual inertial navigation collaborative navigation mechanism is adopted. By simultaneously accessing inertial sensor data and monocular / binocular image streams, an optimized registration is performed based on the forward visual odometry of the reference frame and the IMU trajectory of the current frame. The optimized trajectory is used to construct an "INS-vision" joint state space model to achieve high-frequency positioning updates. The reference trajectory of the historical GNSS high-confidence area is composed of data recorded by multiple missions in the same area. The centerline of the confidence trajectory is extracted using a trajectory clustering algorithm, and the current INS trajectory is constrained and corrected using a Kalman filter. The closed-loop detection mechanism refers to: based on the simultaneous operation of image similarity matching and point cloud geometric reconstruction and comparison, 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.

Images