Real-scene 3D reconstruction method and system

Abstract

This invention belongs to the field of 3D reconstruction technology and discloses a method and system for real-scene 3D reconstruction. The method includes: constructing a training image set covering scenes with and without reflections; training a classifier based on the training image set to efficiently distinguish between photos with and without reflections; for a preset scene, obtaining the drone's shooting points and the corresponding shooting angles for each shooting point through viewpoint planning, performing path planning based on the shooting points, and using the classifier to help the drone acquire modeled photos without reflections; and performing 3D reconstruction of the preset scene based on the modeled photos to obtain a 3D model of the preset scene. This invention detects strong light and reflection problems in images in real time during drone data acquisition, avoiding repeated acquisition of low-quality data and significantly improving the accuracy and efficiency of the entire 3D reconstruction process.

Description

Technical Field

[0001] This invention relates to the field of 3D reconstruction technology, and more specifically, to a method and system for real-scene 3D reconstruction. Background Technology

[0002] In traditional UAV aerial photography used for 3D reconstruction, image quality directly impacts subsequent data processing and reconstruction results. However, because UAVs often operate under complex lighting conditions during actual flight (such as direct sunlight or reflective surfaces), the captured images frequently suffer from overexposure due to strong light or localized highlight reflections. This strong light or reflection not only leads to the loss of image details but also introduces errors in feature point extraction and matching, severely reducing the accuracy and efficiency of 3D reconstruction. Furthermore, to obtain ideal image data and circumvent these problems, traditional methods often rely on favorable weather and soft lighting conditions, such as early morning or dusk. This stringent requirement for external environmental conditions significantly limits the flexibility and time window of UAV operations, resulting in extended project cycles and increased costs, which greatly restricts the flexibility and efficiency of UAV aerial photography.

[0003] Therefore, it is particularly important to develop a technology that can identify images with obvious strong light or reflection problems in real time during the acquisition process.

[0004] In view of this, the present invention proposes a real-scene 3D reconstruction method and system to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and achieve the above objectives, the present invention provides the following technical solution: a real-scene 3D reconstruction method, comprising:

[0006] Construct a training image set covering both reflective and non-reflective scenes;

[0007] Based on the training image set, a classifier is trained to distinguish between photos with and without reflection;

[0008] For a given scenario, viewpoint planning is used to obtain the target area, shooting point, and camera intrinsic and extrinsic parameters for each shot taken by the drone. Path planning is performed based on the shooting point, and a classifier is used to assist the drone in obtaining non-reflective modeled photos.

[0009] Based on the modeling photos, a 3D reconstruction of the preset scene is performed to obtain a 3D model of the preset scene.

[0010] Furthermore, the methods for constructing the training graph set include:

[0011] Collect photos with and without reflections in different scenes; label each photo as either reflective or non-reflective to obtain labeled photos; perform data augmentation on the labeled photos, while keeping the label of a photo unchanged after data augmentation; merge the labeled photos and the data-augmented photos together to obtain a training image set containing M labeled photos with reflections and N labeled photos without reflections, where M and N are of the same order of magnitude.

[0012] Furthermore, the training methods for the classifier include:

[0013] The photos in the training image set are used as input, and the output is a binary label of the reflectivity attribute of the photos. Feature engineering operations are performed on the photos in the training image set to obtain a training dataset. The training dataset is divided into a training set and a validation set, and the selected machine learning model is trained to obtain a classifier for labeling the reflectivity attribute of photos.

[0014] Furthermore, methods for obtaining modeling photos include:

[0015] Step F1F1: The drone takes a picture at the planned shooting point using the planned camera intrinsic and extrinsic parameters. The picture is then cropped to the area corresponding to the target region as a test shot.

[0016] Step F1F2: Use a classifier to determine whether the test photo has reflection. If the test photo has no reflection, use it as the modeling photo and proceed to step F1F4; if the test photo has reflection, proceed to step F1F3.

[0017] Step F1F3: Mark the current shooting point as a reflective shooting point, send feedback to the host computer, and execute the temporary reshoot procedure;

[0018] Step F1F4: The drone automatically moves to the next planned shooting point and repeats step F1F1.

[0019] Furthermore, the steps of the temporary reshoot procedure include:

[0020] Step F2F1: Keep the x-axis direction of the current camera coordinate system unchanged, and determine a spatial unit vector in the world coordinate system that has an angle of less than 90° or greater than 90° with the positive y-axis of the camera coordinate system and is parallel to the short side of the rectangular area as the flight direction of the UAV; define the number of retakes as 0; the rectangular area is a rectangular target area on the plane.

[0021] Step F2: The drone departs from the current shooting point and flies a preset distance s millimeters along the flight direction to reach the temporary shooting point;

[0022] Steps F2 and F3: For the rectangular area, at the temporary shooting point, use the viewpoint planning method to determine the new camera intrinsic and extrinsic parameters so that the captured photo includes the rectangular area.

[0023] Steps F2 and F4: Take a photo, crop the photo to the area corresponding to the rectangle, and use it as a retake photo; increment the retake count by 1.

[0024] In steps F2 and F5, the classifier is used to determine whether the retaken photo has any reflection. If the retaken photo has no reflection, the retaken photo is used as the modeling photo, the current shooting point is marked as the shooting point that has been retaken, feedback is sent to the host computer, and the temporary retake program is terminated.

[0025] If the retaken photo has a reflection, determine whether to repeat steps F2F2 to F2F5 based on the preset total number of times. If steps F2F2 to F2F5 are not repeated, mark the current shooting point as a shooting point that has not been retaken, send feedback to the host computer, and end the temporary retake program.

[0026] Furthermore, if the number of retakes does not exceed the preset total number of times, then steps F2F2 to F2F5 are repeated; if the retaken photo has glare and the number of retakes is equal to the preset total number of times, then the current shooting point is marked as a shooting point that has not been retaken, feedback is sent to the host computer, and the temporary retake program ends.

[0027] Furthermore, different scenarios include: using drones to take photos in different environments, covering different lighting conditions; filtering photos containing reflective properties from existing photo datasets; simulating reflective materials using 3D rendering tools to generate photos with reflective properties; and generating photos containing reflective properties using text-based image models.

[0028] Furthermore, data augmentation methods include: randomly flipping, randomly translating, randomly rotating, adjusting brightness, and simulating motion blur on photos without altering reflective properties.

[0029] A real-scene 3D reconstruction system is used to implement the aforementioned real-scene 3D reconstruction method. The system includes:

[0030] The training preparation module constructs a training image set, covering both reflective and non-reflective scenes.

[0031] The model training module trains a classifier based on the training image set to distinguish between photos with and without reflection.

[0032] The acquisition module, for a preset scene, obtains the target area, shooting point and camera intrinsic and extrinsic parameters corresponding to each shooting point for each shot by the drone through viewpoint planning, performs path planning based on the shooting point, and uses a classifier to assist the drone in obtaining non-reflective modeling photos.

[0033] The 3D reconstruction module performs 3D reconstruction of a preset scene based on the modeling photograph, resulting in a 3D model of the preset scene.

[0034] The beneficial effects of the real-scene 3D reconstruction method and system of the present invention are as follows:

[0035] This invention detects strong light and reflections in images in real time during drone data acquisition, providing timely feedback to the operator and attempting a certain number of automatic retakes. If a reflection-free image of the target area cannot be obtained within a preset number of attempts, the operator will temporarily take over the drone and adjust the shooting position or angle to avoid repeatedly collecting low-quality data and reduce the need for additional flights or retakes. This not only reduces the difficulty and workload of post-processing images but also significantly improves the accuracy and efficiency of the entire 3D reconstruction process.

[0036] Meanwhile, this invention breaks through the traditional limitations of harsh lighting conditions, enabling drones to stably acquire images that meet reconstruction requirements even in complex lighting environments, significantly improving operational efficiency and flexibility of on-site applications. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the real-scene 3D reconstruction system of the present invention;

[0038] Figure 2 This is a schematic diagram of the real-scene 3D reconstruction method of the present invention;

[0039] Figure 3 This is a schematic diagram of a drone-collected and modeled photograph of the present invention under reflective conditions. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Example 1

[0042] See Figures 1-3 This embodiment provides a real-scene 3D reconstruction system, including: a training preparation module, a model training module, an acquisition module, and a 3D reconstruction module.

[0043] The training preparation module constructs a training image set covering both reflective and non-reflective scenes.

[0044] The method for constructing the training image set includes: collecting photos (or images) with reflection (or highlights) and photos (or images) without reflection in different scenes; labeling each photo as having reflection or without reflection to obtain labeled photos; performing data augmentation on the labeled photos, while keeping the label of a photo unchanged after data augmentation; merging the labeled photos and the data-augmented photos together to obtain a training image set containing M labeled photos with reflection and N labeled photos without reflection, where M and N are of the same order of magnitude and the larger the better.

[0045] Different scenarios include: using drones to capture images in various environments (water surfaces, glass curtain walls, metal surfaces, etc.), covering different lighting conditions (strong midday light, low twilight, diffuse reflection on cloudy days, etc.); selecting images containing reflective properties from existing image datasets (such as UIUC dataset, COCO, Flickr Material Database, etc.); simulating reflective materials using 3D rendering tools (such as Blender) to generate images with controllable specular parameters (e.g., generating different intensities of reflective effects by adjusting the material's specular reflection coefficient and ambient occlusion factor); and generating images containing reflective properties using textural graph models (such as StableDiffusion). Real-world camera data covers the complexity of actual scenes, while simulated and synthesized image data supplement rare reflective situations (such as specular highlights at extreme angles).

[0046] Data augmentation methods include random flipping, random translation, random rotation, brightness adjustment, and simulating motion blur (such as applying a Gaussian blur kernel to the image) without altering reflective properties. Data augmentation can improve the robustness of the model to changes in lighting and motion blur, preventing overfitting.

[0047] The model training module trains a classifier based on the training image set to efficiently distinguish between photos with and without reflections.

[0048] The classifier training method includes: taking photos in the training image set as input and outputting binary labels of the photo's reflectivity attributes (i.e., reflective and non-reflective); performing feature engineering operations on the photos in the training image set to obtain a training dataset; dividing the training dataset into a training set and a validation set (the ratio can be set to 8:2), training the selected machine learning model, and finally obtaining a classifier that can reliably label the reflectivity attributes of photos.

[0049] Feature engineering operations include: scaling each image in the training dataset to a preset size (e.g., 512 pixels in both length and width) to obtain images of the preset size; and normalizing the pixel values ​​of the images of the preset size to the range [0,1] (e.g., normalizing the R value of a certain RGB pixel from 128 to 128). G value normalized from 1 to This process yields normalized photos; all normalized photos constitute the training dataset.

[0050] The machine learning model can be a convolutional neural network (CNN) or the lightweight CNN architecture MobileNetV4. If the classifier runs well on the drone, it can be directly deployed on the drone for edge computing; if the classifier requires a high-performance computer, it can be deployed on a high-performance computer, and the photos taken by the drone can be transmitted to the high-performance computer for processing.

[0051] The acquisition module, for a preset scene, obtains the target area (or region of interest), shooting point, and camera intrinsic and extrinsic parameters corresponding to each shooting point for each shot by the drone through viewpoint planning. Based on the shooting points, it performs path planning and uses a classifier to help the drone obtain non-reflective modeled photos.

[0052] Viewpoint planning and path planning for UAVs in 3D reconstruction are existing technologies and will not be elaborated here.

[0053] Methods for obtaining modeling photos include:

[0054] Step F1: The drone takes a picture at the planned shooting point using the planned camera intrinsic and extrinsic parameters. The picture is then cropped to the area corresponding to the target region as a test shot.

[0055] In step F1F2, the classifier determines whether the test photo has any reflection. If the test photo has no reflection, it is used as the modeling photo, and step F1F4 is executed; if the test photo has reflection, step F1F3 is executed.

[0056] Steps F1 and F3 mark the current shooting point as a reflective shooting point, send feedback to the host computer, and execute the temporary reshoot procedure.

[0057] Step F1F4: The drone automatically moves to the next planned shooting point and repeats step F1F1.

[0058] This embodiment uses a rectangular area on a plane as an example to disclose the specific steps of the temporary reshoot procedure. Examples of a planar target area include water surfaces, the exterior walls of buildings, and roads; these target areas can be divided into multiple rectangular areas for separate shooting. At each shooting point, the x-axis of the camera coordinate system is set to be parallel to the long side of the rectangular area; the camera intrinsic and extrinsic parameters must ensure that the captured image contains the rectangular area; if the area corresponding to the captured image is larger than the rectangular area, the captured image should be cropped to the range of the rectangular area. This system sets the shooting point and the origin O of the camera coordinate system (e.g., ...). Figure 3(As shown) The coordinates in the world coordinate system are the same. The camera coordinate system, world coordinate system, camera intrinsic parameters, and camera extrinsic parameters are existing technologies and will not be described in detail here.

[0059] The steps of the temporary reshoot procedure include:

[0060] Step F2F1: Keep the direction of the x-axis of the current camera coordinate system unchanged, such as... Figure 3 As shown, in the world coordinate system, determine a coordinate system whose angle with the positive y-axis of the camera coordinate system is less than 90° or greater than 90°. Figure 3 The flight direction r is defined as the spatial unit vector that is less than 90 degrees from the y-axis of the coordinate system and parallel to the short side of the rectangular region J. The number of retakes is defined as 0, denoted as i = 0. The flight direction exists and is uniquely determined in space. Let the flight direction be r, and ||r|| = 1.

[0061] Step F2: The drone departs from the current shooting point and flies a preset distance s millimeters (e.g., s = 2000 millimeters) along the flight direction to reach the temporary shooting point; Figure 3 As shown, the drone flies along the flight direction determined in steps F2F1, flying s millimeters each time. After multiple flights, it can fly to a shooting position where it can take a photo without reflection. Considering that sunlight is parallel light, the reflected light F of sunlight (S is the incident direction of sunlight) in the rectangular area J on the plane is also parallel light. The reflected light F in the rectangular area J constitutes a spatial range for taking a photo without reflection. As long as the drone flies out of this area, it can take a photo of the rectangular area without reflection.

[0062] Steps F2 and F3, for the rectangular area, at the temporary shooting point, use viewpoint planning to determine the new camera intrinsic and extrinsic parameters of the camera, with the aim of ensuring that the captured photo includes the rectangular area.

[0063] Steps F2 and F4: Take a photo, crop the photo to the area corresponding to the rectangle and use it as a retake photo; increment the number of retakes by 1, denoted as i = i + 1.

[0064] In steps F2F5, a classifier is used to determine whether the retaken photo has any reflection. If the retaken photo has no reflection, it is used as the modeling photo, the current shooting point is marked as a retaken shooting point, feedback is sent to the host computer, and the temporary retake program ends. If the retaken photo has reflection and the number of retakes is less than the preset total number, steps F2F2 to F2F5 are repeated. If the retaken photo has reflection and the number of retakes is equal to the preset total number, the current shooting point is marked as a non-retaken shooting point, feedback is sent to the host computer, and the operator temporarily takes over the drone, adjusts the drone's shooting point and shooting angle through the camera image to obtain a photo without reflection, and ends the temporary retake program.

[0065] For complex shooting scenarios, such as when the target area is a curved surface or an irregular shape, the aforementioned temporary reshoot procedure is too simplistic and cannot achieve the desired technical effect. In this case, the temporary reshoot procedure can be completely replaced by temporary handover by the operator. Manually taking photos is similar to taking pictures with a mobile phone; it's usually easy to find a shooting point and angle that ensures the target area is reflection-free in the photo. Manually taking photos is clearly not impossible for existing technicians. If the operator also cannot find a shooting point and angle to obtain a reflection-free photo, they can take a photo with less reflection as a modeling photo, or mark the current shooting point as an unprocessed reflective shooting point, wait for the lighting conditions to change, and then return to the unprocessed reflective shooting point at a different time to retake the photo.

[0066] During automated drone aerial photography, the system first uses a built-in machine learning model to analyze each captured photo in real time to determine if there are quality issues related to strong light reflection. Once obvious reflection is detected, the system marks the problematic photo on the image acquisition interface, immediately notifies the operator, and initiates a retake procedure. During the retake phase, the drone flies in a fixed direction and adjusts its shooting angle, thus likely flying out of the reflective area in the captured photo. If a reflection-free photo cannot be obtained within a preset number of retakes, the system sends feedback to the operator, who then takes over the drone, adjusts the shooting angle or viewpoint based on the camera feed, and re-acquires an image that meets the requirements. This method of acquiring modeling photos not only effectively avoids tedious post-processing image restoration but also significantly improves the reliability and accuracy of data during 3D reconstruction. The entire process achieves seamless integration of automatic detection, automatic retake, and temporary operator intervention, fully leveraging the advantages of efficient drone data acquisition and intelligent machine learning judgment. This results in time savings, cost reduction, and improved modeling effectiveness in engineering practice.

[0067] The 3D reconstruction module performs 3D reconstruction of a preset scene based on the modeling photograph, resulting in a 3D model of the preset scene.

[0068] The 3D reconstruction process includes: preprocessing the modeling photograph, which includes noise reduction, contrast enhancement, and distortion correction; establishing the correspondence between points in the 2D photograph and points in 3D space using the camera's intrinsic and extrinsic parameters; and performing 3D reconstruction using methods such as the 3D Gaussian splashing method to obtain a 3D model of the preset scene. The 3D reconstruction process from the modeling photograph to the 3D model is existing technology and will not be elaborated upon here.

[0069] If a high level of visual effect is required, the 3D Gaussian splatter method is recommended (intuitively, many colored ink droplets are suspended in 3D space, and these colored ink droplets are distributed and superimposed in space, which looks consistent with the real 3D scene); if a high level of measurability is required, the oblique photogrammetry 3D reconstruction method is recommended (including SfM-MVS process, meshing, texture mapping and other steps).

[0070] This embodiment proposes a real-scene 3D reconstruction system, which effectively solves the problem of 3D reconstruction caused by strong light and reflection in the existing technology during drone aerial photography.

[0071] Specifically, this system provides a solution for automatic identification, automatic re-taking, and real-time feedback.

[0072] Compared with existing technologies, this system has significant advantages:

[0073] This system, based on machine learning technology, utilizes a model trained on a large number of samples to quickly identify glare and reflection issues in each frame of an image and provide timely feedback to the operator. If, after a certain number of automatic retakes, a reflection-free image still cannot be obtained, the system will provide feedback to the operator, who will then manually adjust the shooting point and angle to obtain a reflection-free image, thus significantly reducing the acquisition of low-quality photos. Reflection-free images provide clearer data, supporting more accurate feature detection. For example, in highly reflective scenes such as glass curtain walls or water surfaces, traditional methods may result in incomplete models due to missing data, while this system can stably output high-quality 3D models.

[0074] This system significantly reduces the time and labor costs of drone operations by minimizing the workload of repeated flights and post-processing image restoration. For example, traditional methods may require multiple return flights for adjustments, while this system minimizes the number of return flights through automatic re-shooting and real-time feedback.

[0075] Therefore, this system not only effectively reduces feature extraction and matching errors caused by strong light or reflection during subsequent 3D reconstruction, but also shortens the overall cycle of data preprocessing and 3D reconstruction, improving the accuracy and completeness of the final model. At the same time, this system overcomes the limitations of traditional methods in terms of demanding lighting conditions, providing technical support for the application of UAVs in a wider range of scenarios.

[0076] In urban planning, this system can be used to quickly generate high-precision 3D models of building facades, maintaining stable performance even under direct sunlight, thus shortening project cycles.

[0077] In the protection of cultural heritage, this invention can efficiently collect image data of ancient buildings, avoid the loss of details due to lighting problems, and ensure the authenticity and integrity of the digitization.

[0078] Example 2

[0079] See Figure 2 As shown, this embodiment provides a real-scene 3D reconstruction method, including:

[0080] Construct a training image set covering both reflective and non-reflective scenes;

[0081] Based on the training image set, a classifier is trained to efficiently distinguish between photos with and without reflection.

[0082] For a given scenario, viewpoint planning is used to obtain the target area, shooting point, and camera intrinsic and extrinsic parameters for each shot taken by the drone. Path planning is performed based on the shooting point, and a classifier is used to assist the drone in obtaining non-reflective modeled photos.

[0083] Based on the modeling photos, a 3D reconstruction of the preset scene is performed to obtain a 3D model of the preset scene.

[0084] For any parts not mentioned in this application, existing technologies may be used or referenced.

[0085] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0086] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A real-scene 3D reconstruction method, characterized in that, include: Construct a training image set covering both reflective and non-reflective scenes; Based on the training image set, a classifier is trained to distinguish between photos with and without reflection; For a given scenario, viewpoint planning is used to obtain the target area, shooting point, and camera intrinsic and extrinsic parameters for each shot taken by the drone. Path planning is then performed based on the shooting point, and a classifier is used to assist the drone in obtaining non-reflective modeled photos. Methods for obtaining modeling photos include: Step F1F1: The drone takes a picture at the planned shooting point using the planned camera intrinsic and extrinsic parameters. The picture is then cropped to the area corresponding to the target region as a test shot. Step F1F2: Use a classifier to determine whether the test photo has reflection. If the test photo has no reflection, use it as the modeling photo and proceed to step F1F4; if the test photo has reflection, proceed to step F1F3. Step F1F3: Mark the current shooting point as a reflective shooting point, send feedback to the host computer, and execute the temporary reshoot procedure; Step F1F4: The drone automatically moves to the next planned shooting point and repeats step F1F1. The steps of the temporary reshoot procedure include: Step F2F1, maintain the current camera coordinate system With the axis direction unchanged, determine an coordinate system in the world coordinate system that is perpendicular to the camera coordinate system. The included angle in the positive direction of the axis is less than or greater than The spatial unit vector parallel to the short side of the rectangular region is used as the flight direction of the UAV; the number of retakes is defined as 0; the rectangular region is a rectangular target area on the plane; Step F2: The drone departs from the current shooting point and flies a preset distance along the flight direction. Millimeters, arrived at the temporary filming location; Steps F2 and F3: For the rectangular area, at the temporary shooting point, use the viewpoint planning method to determine the new camera intrinsic and extrinsic parameters so that the captured photo includes the rectangular area. Steps F2 and F4: Take a photo, crop the photo to the area corresponding to the rectangle, and use it as a retake photo; increment the retake count by 1. In steps F2 and F5, the classifier is used to determine whether the retaken photo has any reflection. If the retaken photo has no reflection, the retaken photo is used as the modeling photo, the current shooting point is marked as the shooting point that has been retaken, feedback is sent to the host computer, and the temporary retake program is terminated. If the retaken photo has a reflection, determine whether to repeat steps F2F2 to F2F5 based on the preset total number of times. If steps F2F2 to F2F5 are not repeated, mark the current shooting point as a shooting point that has not been retaken, send feedback to the host computer, and end the temporary retake program. Based on the modeling photos, a 3D reconstruction of the preset scene is performed to obtain a 3D model of the preset scene.

2. The real-scene 3D reconstruction method according to claim 1, characterized in that, Methods for constructing training graph sets include: Collect photos with and without reflections from different scenes; label each photo as either reflective or non-reflective, obtaining labeled photos; perform data augmentation on the labeled photos, keeping the label unchanged after data augmentation; merge the labeled photos and the data-augmented photos together to obtain a composite image containing... Zhang has already marked the reflective photos and The training image set of Zhang's labeled, non-reflective photographs. and They are on the same order of magnitude.

3. The real-scene 3D reconstruction method according to claim 1, characterized in that, Classifier training methods include: The photos in the training image set are used as input, and the output is a binary label of the reflectivity attribute of the photos. Feature engineering operations are performed on the photos in the training image set to obtain a training dataset. The training dataset is divided into a training set and a validation set, and the selected machine learning model is trained to obtain a classifier for labeling the reflectivity attribute of photos.

4. The real-scene 3D reconstruction method according to claim 1, characterized in that, If the number of retakes does not exceed the preset total number of times, then repeat steps F2F2 to F2F5; if the retaken photo has glare and the number of retakes is equal to the preset total number of times, then mark the current shooting point as a shooting point that has not been retaken, send feedback to the host computer, and end the temporary retake program.

5. The real-scene 3D reconstruction method according to claim 2, characterized in that, Different scenarios include: using drones to take photos in different environments, covering different lighting conditions; filtering photos containing reflective properties from existing photo datasets; simulating reflective materials using 3D rendering tools to generate photos with reflective properties; and generating photos containing reflective properties using text-based image models.

6. The real-scene 3D reconstruction method according to claim 2, characterized in that, Data augmentation methods include randomly flipping, randomly translating, randomly rotating, adjusting brightness, and simulating motion blur on photos without altering reflective properties.

7. A real-scene 3D reconstruction system, characterized in that, The system for implementing the real-scene 3D reconstruction method according to any one of claims 1-6 includes: The training preparation module constructs a training image set, covering both reflective and non-reflective scenes. The model training module trains a classifier based on the training image set to distinguish between photos with and without reflection. The acquisition module, for a preset scene, obtains the target area, shooting point and camera intrinsic and extrinsic parameters corresponding to each shooting point for each shot by the drone through viewpoint planning, performs path planning based on the shooting point, and uses a classifier to assist the drone in obtaining non-reflective modeling photos. The 3D reconstruction module performs 3D reconstruction of a preset scene based on the modeling photograph, resulting in a 3D model of the preset scene.

Images