Unmanned aerial vehicle path planning method based on bayesian reconstruction and improved grey wolf algorithm
By combining Bayesian reconstruction and an improved Grey Wolf algorithm, targeted reshoot routes are generated, solving the problems of uncertainty ignoring and low optimization efficiency in UAV 3D reconstruction, and improving the geometric accuracy and texture fidelity of 3D reconstruction.
Patent Information
- Application Number
- CN202611016163.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-09
- Publication Date
- 2026-08-25
AI Technical Summary
Existing UAV 3D reconstruction methods ignore reconstruction uncertainty information, lack targeted reshoot perspectives, and standard intelligent optimization algorithms are inefficient and prone to getting trapped in local optima, resulting in holes, geometric distortion, or blurred textures in the 3D reconstruction model.
Bayesian reconstruction is used to extract uncertainty information, and the camera angle is optimized by combining the improved Grey Wolf algorithm to generate targeted waypoints for reshooting. The UAV is then controlled through closed loop to perform reshooting and update the 3D reconstruction model.
It significantly improves the geometric accuracy and texture fidelity of 3D reconstruction, avoids the problems of blind reshooting and low efficiency of traditional methods, and achieves efficient focusing on weak areas of reconstruction.
Smart Images
Figure CN122632853A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision and UAV path planning technology, specifically relating to a UAV path planning method based on Bayesian reconstruction and an improved gray wolf algorithm. Background Technology
[0002] Currently, 3D reconstruction using images acquired by drones has been widely applied in fields such as bridge inspection, urban modeling, and cultural heritage protection. However, the quality of 3D reconstruction is highly dependent on the viewing angle and coverage integrity of the image acquisition. While traditional preset flight paths can quickly obtain an overview of the scene, they are difficult to achieve optimal coverage for complex structures, resulting in holes, geometric distortion, or blurred textures in the reconstructed model.
[0003] Existing path planning methods based on initial coarse models establish a rough model through the first flight and then plan the path for the second reshoot based on this model. These methods can be used offline and are compatible with commercial drones, but most ignore uncertainties in the 3D reconstruction process and cannot actively identify and focus on weak areas in the reconstruction. Meanwhile, standard intelligent optimization algorithms have low search efficiency in high-dimensional viewpoint planning problems, are prone to getting trapped in local optima, and typically require numerous iterations to converge.
[0004] In recent years, 3D Gaussian splashing and its Bayesian extension techniques have enabled high-fidelity, real-time rendering of 3D reconstructions, outputting the posterior variance of each Gaussian point as a measure of reconstruction uncertainty. However, there is currently no mature solution for effectively integrating this uncertainty information into reshoot path planning to form a closed loop of "uncertainty modeling—path planning—reshoot execution—quality assessment." Therefore, there is an urgent need for a method that can utilize Bayesian reconstruction of uncertainty and combine it with efficient optimization algorithms to generate targeted reshoot paths, thereby improving the geometric accuracy and texture fidelity of 3D reconstructions. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a UAV path planning method based on Bayesian reconstruction and an improved gray wolf algorithm, which solves the problems of ignoring reconstruction uncertainty, lack of targeted reshoot perspectives, low efficiency of optimization algorithms, and easy getting trapped in local optima in existing technologies.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a UAV path planning method based on Bayesian reconstruction and an improved gray wolf algorithm, comprising the following steps:
[0007] S1. Construct a 3D model of the target object based on Bayesian 3D reconstruction, and extract the uncertainty information of the 3D model to generate an uncertainty point cloud; S2. Constrain the camera position on a fixed-height horizontal ring around the target object, and use a one-dimensional improved gray wolf algorithm to optimize the camera's horizontal angle. With the goal of maximizing the sum of the visible uncertainties of the candidate viewpoints, generate waypoints for reshooting. S3. Convert the waypoints for the reshoot into geographic coordinates, generate an executable flight path file for the drone, control the drone to perform the reshoot, and integrate the new images to update the 3D reconstruction model.
[0008] Preferably, step S1 specifically includes: A multi-view image sequence of the target object is collected by UAV, and Bayesian 3D reconstruction is performed using the Horseshoe Splatting method. The variance of the scale parameter of each Gaussian point is calculated by posterior sampling, and this variance is used as an uncertainty measure to form a Gaussian point cloud with uncertainty attributes. The Horseshoe Splatting method employs global-local sparse priors and variational inference for model training.
[0009] Preferably, step S2 further includes: By using GPS information from photos and the COLMAP camera center to solve the 3D similarity transformation, the uncertain point cloud is transformed from the local coordinate system to the UTM planar coordinate system.
[0010] Preferably, the improved gray wolf algorithm includes: The three-dimensional viewpoint search space is reduced to a one-dimensional horizontal angle variable, and the horizontal annulus is divided into multiple sectors. Gray Wolf optimization is performed independently in each sector. The improved gray wolf algorithm employs a nonlinear convergence factor, the expression of which is:
[0011] in This represents the current iteration number. This represents the maximum number of iterations.
[0012] Preferably, the center of the horizontal ring in step S2 is the ground projection point of the bottom center of the target object, the radius is set according to the target size and the safe distance, and the flight altitude is set according to the target height and the safe clearance.
[0013] Preferably, the fitness function of the improved gray wolf algorithm is: The sum of uncertainties of all Gaussian points visible from the candidate viewpoint; if the distance between the candidate camera position and the point cloud is less than the safety threshold, the fitness value is set to 0.
[0014] Preferably, step S3 includes: Convert the waypoints in the UTM coordinate system to WGS84 latitude and longitude coordinates to generate a KMZ route file that conforms to UAV specifications. The file includes waypoint latitude and longitude, altitude, flight speed, and turning mode.
[0015] Preferably, step S3 also outputs the nose direction angle and gimbal pitch angle for each waypoint, wherein the nose direction points to the center of the horizontal circle.
[0016] Preferably, after the reshoot is completed, the newly acquired image is merged with the original image, the 3D reconstruction model is retrained, and the reshoot effect is quantified by at least one of the following indicators: total uncertainty, reprojection error, rendering quality, Gaussian ellipsoid consistency, coverage uniformity, and color smoothness.
[0017] Preferably, in step S2, the number of sectors in the horizontal ring is 8, 12, or 16, the number of gray wolf optimized populations in each sector is 20, and the number of iterations is 30.
[0018] Compared with existing technologies, this invention provides a UAV path planning method based on Bayesian reconstruction and an improved gray wolf algorithm, which has the following beneficial effects: This invention is the first to introduce posterior uncertainty information from Bayesian reconstruction into UAV reshooting path planning. By extracting the posterior variance of each Gaussian point through Horseshoe Splatting as an uncertainty map, the reshooting perspective can be accurately focused on the weak areas of reconstruction, overcoming the shortcomings of traditional methods such as blind shooting and low reshooting efficiency.
[0019] This invention addresses the characteristics of structured targets such as buildings by reducing the viewpoint optimization problem in three-dimensional space to a one-dimensional angle variable. It also employs an improved Grey Wolf algorithm with independent optimization of circular sectors, which significantly reduces the search space complexity, improves the convergence speed and optimization quality, and effectively avoids the problem of high-dimensional optimization easily getting trapped in local optima.
[0020] This invention achieves a complete automated closed loop from planning to execution through coordinate transformation and simplified KMZ file generation. The generated flight path can be directly imported into commercial drones, solving the "last mile" problem between algorithm research and actual flight.
[0021] This invention constructs a multi-dimensional quantitative evaluation system for reshoot effects, including indicators such as total uncertainty, reprojection error, rendering quality, Gaussian ellipsoid geometric consistency, distribution uniformity, and color smoothness, providing an objective and comprehensive evaluation basis for the effectiveness of reshoot strategies. Attached Figure Description
[0022] Figure 1 This is an overall flowchart of the UAV 3D reconstruction path planning method based on Bayesian reconstruction and improved gray wolf algorithm proposed in this invention; Figure 2 This is a schematic diagram of the improved gray wolf algorithm supplementary shooting path planning process in step S2 of the present invention; Figure 3 This is a schematic diagram of collecting data from an ancient pagoda in an embodiment of the present invention; Figure 4 This is an example of the uncertainty point cloud of the ancient pagoda in this invention embodiment; Figure 5 This is the image of the ancient pagoda rendered for the first time by the algorithm in this embodiment of the invention; Figure 6 This is a schematic diagram of waypoint planning in the UTM coordinate system in an embodiment of the present invention; Figure 7 This is an image of the ancient pagoda rendered by the algorithm after reshooting in an embodiment of the present invention. Detailed Implementation
[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0024] See attached document Figure 1 and Figure 2 The UAV path planning method based on Bayesian reconstruction and improved Grey Wolf algorithm includes the following steps: S1. Construct a 3D model of the target object based on Bayesian 3D reconstruction, and extract the uncertainty information of the 3D model to generate an uncertainty point cloud; This step specifically includes: A multi-view image sequence of the target object is collected by UAV, and Bayesian 3D reconstruction is performed using the Horseshoe Splatting method. The variance of the scale parameter of each Gaussian point is calculated by posterior sampling, and this variance is used as an uncertainty measure to form a Gaussian point cloud with uncertainty attributes. The Horseshoe Splatting method employs global-local sparse priors and variational inference for model training. S2. Constrain the camera position on a fixed-height horizontal ring around the target object, and use a one-dimensional improved gray wolf algorithm to optimize the camera's horizontal angle. With the goal of maximizing the sum of the visible uncertainties of the candidate viewpoints, generate waypoints for reshooting. This step also includes: By using GPS information from photos and the COLMAP camera center to solve the 3D similarity transformation, the uncertain point cloud is transformed from the local coordinate system to the UTM planar coordinate system; The improved Grey Wolf algorithm includes: The three-dimensional viewpoint search space is reduced to a one-dimensional horizontal angle variable, and the horizontal annulus is divided into multiple sectors. Gray Wolf optimization is performed independently in each sector. The improved gray wolf algorithm employs a nonlinear convergence factor, the expression of which is:
[0025] in This represents the current iteration number. This represents the maximum number of iterations.
[0026] The center of the horizontal ring in this step is the ground projection point of the bottom center of the target object, the radius is set according to the target size and safe distance, and the flight altitude is set according to the target height and safe clearance.
[0027] The fitness function of the improved gray wolf algorithm is: The sum of uncertainties of all Gaussian points visible from the candidate viewpoint; if the distance between the candidate camera position and the point cloud is less than the safety threshold, the fitness value is set to 0. In step S2, the number of sectors in the horizontal ring is 8, 12, or 16, the number of gray wolf optimized populations in each sector is 20, and the number of iterations is 30. S3. Convert the waypoints for the reshoot into geographic coordinates, generate an executable flight path file for the drone, control the drone to perform the reshoot, and integrate the new images to update the 3D reconstruction model.
[0028] This step includes: Convert the waypoints in the UTM coordinate system to WGS84 latitude and longitude coordinates to generate a KMZ route file that conforms to UAV specifications. The file includes waypoint latitude and longitude, altitude, flight speed, and turning mode.
[0029] This step also outputs the nose direction angle and gimbal pitch angle for each waypoint, where the nose direction points to the center of the horizontal circle.
[0030] In this step, after the reshoot is completed, the newly acquired image is merged with the original image, the 3D reconstruction model is retrained, and the reshoot effect is quantified by at least one of the following indicators: total uncertainty, reprojection error, rendering quality, Gaussian ellipsoid consistency, coverage uniformity, and color smoothness.
[0031] This invention addresses the problems of existing UAV 3D reconstruction path planning methods neglecting 3D reconstruction uncertainty information and lacking specificity in reshoot perspectives, as well as the shortcomings of standard intelligent optimization algorithms in high-dimensional viewpoint planning such as low search efficiency and easy getting trapped in local optima. By deeply coupling the posterior uncertainty metric in Bayesian reconstruction with the improved Grey Wolf algorithm, this invention innovatively realizes uncertainty-driven reshoot path planning.
[0032] This method, while ensuring compatibility with commercial drones and enabling low-cost deployment, generates efficient reshoot routes focusing on weak areas of reconstruction, providing a highly efficient technical means for building 3D reconstruction, digital twins, and intelligent operation and maintenance. The following description, in conjunction with accompanying figures and specific examples, illustrates the drone 3D reconstruction path planning method based on Bayesian reconstruction and an improved Grey Wolf algorithm proposed in this invention.
[0033] Example: Path planning and 3D reconstruction quality improvement for reshooting an octagonal pavilion-style ancient pagoda See attached document Figures 3 to 7 This embodiment uses an octagonal, nine-story pavilion-style ancient pagoda as the subject. This pagoda is a typical Ming and Qing dynasty Jiangnan-style brick and stone pagoda, exhibiting a variable cross-section. Each story features eaves, brackets, and doors and windows. The finial is made of metal, while the main body is constructed of whitewashed brick and stone, with localized weathering textures and variations in reflectivity. The densely packed details of the eaves and brackets place high demands on the completeness of detail and geometric accuracy in the 3D reconstruction. This embodiment verifies the effectiveness of the method of this invention on medium-scale, complex traditional architectural structures. Figure 3 The image shown is a schematic diagram of the ancient pagoda data collection in this embodiment. The specific steps include: S1. Construction and Uncertainty Extraction of 3D Models Based on Bayesian Reconstruction S11. Multi-view image acquisition and initial reconstruction Using a DJI Air 3S drone, a horizontal ring at a fixed height was established around the ancient pagoda: the center was taken as the geometric center of the pagoda's base, the radius of the ring was determined to be 15 meters based on the pagoda's base diameter plus a safety margin, and the flight altitude was set to 25 meters. The drone flew along this horizontal ring at a constant speed, hovering and taking pictures every 30°. Simultaneously, a waypoint tweening was inserted between every two adjacent waypoints on the ring, resulting in 24 waypoints and 24 original images. During the flight, the GPS position and gimbal attitude were recorded for each photo. After acquisition, the image sequence was input into COLMAP software for sparse reconstruction, obtaining the camera pose and initial sparse point cloud. S12, Horseshoe Splatting Bayesian 3D Reconstruction The image sequence and camera pose are input into the Horseshoe Splatting framework for training. Global-local sparse prior and variational inference are employed. After training, the scale parameter of each Gaussian point is posteriorly sampled, and the variance of the scale parameter for each Gaussian point is calculated. This variance is used as the uncertainty measure for that point. Finally, a dense Gaussian point cloud with uncertainty properties is generated, denoted as:
[0034] in For location, Let covariance matrix be the variance matrix. For opacity, This represents the posterior variance (the value of uncertainty). For example... Figure 4 The figure shown is an uncertain point cloud of the ancient pagoda in an embodiment of the present invention; S2. UTM coordinate system reshooting path planning based on the improved Grey Wolf algorithm S21. Coordinate Transformation and Uncertain Point Cloud Transformation Using the GPS coordinates from the first-stage flight photos and the camera center coordinates output by COLMAP, a 3D similarity transformation is solved to transform the Gaussian point cloud generated in step S12 from the local coordinate system to the UTM plane coordinate system. Simultaneously, the GPS average altitude from the first-stage flight photos (…) Subtract the drone's set altitude from the ground to calculate the ground elevation. The converted UTM point cloud is denoted as... ; S22, Fixed Height Horizontal Ring Parameter Setting In this embodiment, the center and radius of the fixed-height horizontal ring are pre-specified manually: the center is taken as the ground projection point of the geometric center of the bottom of the ancient pagoda, denoted as... The radius is set according to the tower size and safety distance. Meters. The camera's flight altitude remains consistent with the first phase, still at [missing information]. rice; S23, Improved Grey Wolf Algorithm for Independent Optimization of Circular Sectors Because the camera position is determined only by the horizontal angle The decision was made to reduce the 3D search space to a 1D angle variable. The circumference was divided into 12 equal sectors, and the Grey Wolf optimization algorithm was run independently in each sector. The population size was 20, and the number of iterations was 30. The fitness function was defined as: given candidate horizontal angles... Corresponding camera position
[0035] The camera is always pointing towards the center of the circle. The gimbal pitch angle remains the same as in the first stage to cover the middle to the top of the tower. Calculate the sum of uncertainties for all observable Gaussian points from this viewpoint:
[0036] Visibility assessment employs an approximate intersection test between a ray and a Gaussian ellipsoid, taking into account self-occlusion by the tower. Collision detection is also included: if the nearest distance between the candidate camera position and the UTM point cloud is less than a safety threshold... Rice, then Set it directly to 0. The nonlinear convergence factor is adopted. ,in This represents the current iteration number. ; After independent optimization within each sector, 12 locally optimal angles were obtained. The corresponding camera positions are the waypoints for reshooting. Since the first stage already collected 12 uniform angles (0°, 30°, 60°, ..., 330°) and encrypted waypoints every 15°, the optimal angle output by this algorithm may fall near existing waypoints or at different locations. To avoid duplicate shooting, a minimum angle interval threshold is set. If the angle difference between a reshoot waypoint and an existing waypoint is less than the threshold, the waypoint is discarded, and a suboptimal angle is selected within the sector. Ultimately, eight valid reshoot waypoints were generated. Figure 2 This example illustrates the circular and sector divisions, as well as the optimized distribution of waypoints for re-enhancing images. Red dots represent waypoints output by the algorithm, while gray dots represent waypoints already acquired in the first stage. Figure 6 The diagram shown is a schematic diagram of waypoint planning in the UTM coordinate system in an embodiment of the present invention, where the red dots represent the waypoints for reshooting output by the algorithm. S3, Coordinate Transformation and UAV Flight Path Generation S31. Generate KMZ route file Convert the 8 UTM waypoints obtained in step S23 back to WGS84 latitude and longitude coordinates. Use a simplified KMZ format, containing only the waypoint latitude and longitude, altitude, flight speed, and turn mode. Set the nose-up angle of each waypoint to the azimuth angle from the current waypoint to the center of the circle, and the gimbal pitch angle to -20°. Write these parameters into the pagoda_waypoints.kmz file, which can be directly imported into the UAV remote controller.
[0037] S32, Perform reshoot flight and model update Upload the generated KMZ flight path file to the drone and perform a reshoot flight. After the reshoot flight is completed, merge the 8 newly acquired images with the original 24 images, and retrain the Horseshoe Splatting model with the same training parameters as in step S12.
[0038] S4. Quantitative Evaluation of Reshoot Results To objectively evaluate the reshooting effect of this invention, the following six indicators were calculated before and after the reshooting, and the algorithm-rendered image before the reshooting (e.g.) was compared with the image before the reshooting. Figure 5 (as shown) and the re-shot rendered image (as shown) Figure 7 Compare the results shown below: (1) Total uncertainty: The sum of the posterior variances of all Gaussian points before the reshoot was 156.7, which dropped to 51.2 after the reshoot, a relative decrease of 67.3%, indicating that the weak areas of reconstruction were effectively strengthened.
[0039] (2) Average reprojection error: The error of 3D Gaussian projection onto the image plane is calculated on the training set image. Before the reshoot, it was 2.15 pixels, and after the reshoot, it dropped to 0.84 pixels, a reduction of 60.9%, indicating that the geometric accuracy was significantly improved.
[0040] (3) Rendering quality: The peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) of the rendered image and the real image were calculated on 8 reserved test images. Before the reshoot, the PSNR was 23.1 dB and the SSIM was 0.789. After the reshoot, the PSNR increased to 30.8 dB and the SSIM increased to 0.921, and the texture fidelity and structural consistency were significantly improved. Figure 5 This is the image of the ancient pagoda rendered for the first time by the algorithm in this embodiment of the invention. Figure 7 The image of the ancient pagoda rendered by the algorithm after reshooting is visible. Figure 7 Compare Figure 5 It is significantly clearer.
[0041] (4) Gaussian ellipsoid scale / direction consistency: The standard deviation of the angle between the principal scale directions of adjacent Gaussian points decreased from 21.3° before the reshoot to 10.5° after the reshoot, a decrease of 50.7%. The Gaussian ellipsoids are more regular and the local geometric smoothness is improved.
[0042] (5) Uniformity of Gaussian sphere coverage density: The reconstructed space was divided into voxel grids with a side length of 0.2 meters. The number of Gaussian spheres in each voxel was counted, and the coefficient of variation (CV) of the number of Gaussian spheres in all non-empty voxels was calculated. The CV before the reshoot was 0.68, and the CV after the reshoot dropped to 0.39, a decrease of 42.6%, indicating that the distribution of Gaussian spheres was more uniform, the obstruction area under the eaves and other areas no longer had the phenomenon of accumulation, and the sparse areas such as the back side of the bracket were effectively filled.
[0043] (6) Color smoothness: The average color difference between adjacent Gaussian points decreased from 0.095 (RGB difference) before the reshoot to 0.044 after the reshoot, a decrease of 53.7%, resulting in a more natural texture transition and significantly enhanced realism of weathering texture.
[0044] The above results demonstrate that the total uncertainty, reprojection error, rendering quality, geometric consistency, distribution uniformity, and color smoothness of the model are significantly improved after reshooting. This embodiment successfully verifies the effectiveness of the method of the present invention in complex traditional brick and stone ancient pagoda scenes: by using only a single fixed-height horizontal circular flight path, extracting the uncertainty map through Bayesian reconstruction, and then optimizing the reshoot angle using a one-dimensional improved Grey Wolf algorithm, the detail integrity and geometric accuracy of the 3D reconstruction can be significantly improved without increasing the flight altitude layer.
[0045] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A UAV path planning method based on Bayesian reconstruction and improved gray wolf algorithm, characterized in that, Includes the following steps: S1. Construct a 3D model of the target object based on Bayesian 3D reconstruction, and extract the uncertainty information of the 3D model to generate an uncertainty point cloud; S2. Constrain the camera position on a fixed-height horizontal ring around the target object, and use a one-dimensional improved gray wolf algorithm to optimize the camera's horizontal angle. With the goal of maximizing the sum of the visible uncertainties of the candidate viewpoints, generate waypoints for reshooting. S3. Convert the waypoints for the reshoot into geographic coordinates, generate an executable flight path file for the drone, control the drone to perform the reshoot, and integrate the new images to update the 3D reconstruction model.
2. The UAV path planning method based on Bayesian reconstruction and improved gray wolf algorithm according to claim 1, characterized in that, Step S1 specifically includes: A multi-view image sequence of the target object is collected by UAV, and Bayesian 3D reconstruction is performed using the Horseshoe Splatting method. The variance of the scale parameter of each Gaussian point is calculated by posterior sampling, and this variance is used as an uncertainty measure to form a Gaussian point cloud with uncertainty attributes. The Horseshoe Splatting method employs global-local sparse priors and variational inference for model training.
3. The UAV path planning method based on Bayesian reconstruction and improved gray wolf algorithm according to claim 1, characterized in that, Step S2 also includes: By using GPS information from photos and the COLMAP camera center to solve the 3D similarity transformation, the uncertain point cloud is transformed from the local coordinate system to the UTM planar coordinate system.
4. The UAV path planning method based on Bayesian reconstruction and improved gray wolf algorithm according to claim 1, characterized in that, The improved Grey Wolf algorithm includes: The three-dimensional viewpoint search space is reduced to a one-dimensional horizontal angle variable, and the horizontal annulus is divided into multiple sectors. Gray Wolf optimization is performed independently in each sector. The improved gray wolf algorithm employs a nonlinear convergence factor, the expression of which is: in This represents the current iteration number. This represents the maximum number of iterations.
5. The UAV path planning method based on Bayesian reconstruction and improved gray wolf algorithm according to claim 1, characterized in that, In step S2, the center of the horizontal ring is the ground projection point of the bottom center of the target object, the radius is set according to the target size and the safe distance, and the flight altitude is set according to the target height and the safe clearance.
6. The UAV path planning method based on Bayesian reconstruction and improved gray wolf algorithm according to claim 1, characterized in that, The fitness function of the improved gray wolf algorithm is: The sum of uncertainties of all Gaussian points visible from the candidate viewpoint; if the distance between the candidate camera position and the point cloud is less than the safety threshold, the fitness value is set to 0.
7. The UAV path planning method based on Bayesian reconstruction and improved gray wolf algorithm according to claim 1, characterized in that, Step S3 includes: Convert the waypoints in the UTM coordinate system to WGS84 latitude and longitude coordinates to generate a KMZ flight path file that conforms to UAV specifications. The file includes waypoint latitude and longitude, altitude, flight speed, and turning mode.
8. The UAV path planning method based on Bayesian reconstruction and improved gray wolf algorithm according to claim 7, characterized in that, Step S3 also outputs the nose direction angle and gimbal pitch angle for each waypoint, where the nose direction points to the center of the horizontal circle.
9. The UAV path planning method based on Bayesian reconstruction and improved gray wolf algorithm according to claim 1, characterized in that, After the reshoot is completed, the newly acquired image is merged with the original image, the 3D reconstruction model is retrained, and the reshoot effect is quantified by at least one of the following indicators: total uncertainty, reprojection error, rendering quality, Gaussian ellipsoid consistency, coverage uniformity, and color smoothness.
10. The UAV path planning method based on Bayesian reconstruction and improved gray wolf algorithm according to claim 4, characterized in that, In step S2, the number of sectors in the horizontal ring is 8, 12, or 16, the number of gray wolf optimized populations in each sector is 20, and the number of iterations is 30.