Three-dimensional forest reconstruction method based on 3DGS technology

Abstract

The invention discloses a three-dimensional forest reconstruction method based on a 3DGS technology. The method comprises the following steps: step 1, obtaining a picture through a lateral overlap ratio video stream image interception algorithm based on a threshold value; 2, carrying out image feature extraction, matching and sparse reconstruction to obtain a data set comprising sparse point cloud and camera poses; step 3, introducing forest three-dimensional model reconstruction based on a 3DGS technology; and 4, carrying out visual rendering on the model, and displaying an interactive three-dimensional scene. According to the threshold-based lateral overlapping ratio video stream image interception algorithm, in combination with key information such as an image overlapping ratio, videos shot by an unmanned aerial vehicle are generated into a high-quality image set; the 3DGS technology is applied to three-dimensional forest reconstruction, complex structures such as trees and terrains in the forest are accurately expressed through Gaussian distribution, reconstruction errors are effectively reduced, rendering quality is improved, and finally man-machine interaction is achieved through presentation of a visual interface.

Description

Technical Field

[0001] The invention relates to a new forest monitoring method for three-dimensional reconstruction based on 3DGS (3D-Gaussian-Splatting) technology, and belongs to the field of forest monitoring. Background Art

[0002] Forests are one of the most important ecosystems on Earth. They not only maintain the global ecological balance, but also have multiple ecological functions such as protecting species and fixing carbon and releasing oxygen. Among them, forest monitoring plays a particularly important role, mainly including monitoring of forest coverage, forest productivity, and forest health. By providing real-time and accurate data, it provides a scientific basis for forest resource protection and management, and is the key to ensuring the health, stability and sustainable development of forest ecosystems.

[0003] At present, the methods used for forest monitoring mainly include continuous forest resource inventory, remote sensing technology, and drone monitoring technology. Continuous forest resource inventory is to establish a regression relationship between initial inspection and re-inspection by regularly repeating observations of fixed sample plots to estimate the current status and dynamic changes of forest resources; remote sensing technology uses satellites or drones and other remote sensing platforms to obtain information, and combines ground survey data to establish models for estimation; drone monitoring technology is used for tree species identification, crown width estimation, breast diameter estimation, etc., combined with geographic information system technology to improve monitoring accuracy.

[0004] However, the above forest monitoring methods still have some shortcomings. For example, the continuous inventory of forest resources adopts fixed sample plot surveys, which is a relatively primitive sampling method, with a large workload and low efficiency; remote sensing technology is difficult to construct a high-precision three-dimensional scene model due to the limitation of spatial resolution, that is, the resolution is low; although drone monitoring can improve accuracy to a certain extent, the flight time and load are limited. In addition, remote sensing technology and drone monitoring technology can only obtain planar information of the forest and cannot accurately express the three-dimensional spatial relationship between trees.

[0005] Among them, three-dimensional reconstruction technology can convert two-dimensional information into a three-dimensional model, which can better show the three-dimensional spatial relationship. In 2014, Zhang Tao from Xidian University proposed a three-dimensional reconstruction method based on monocular vision. This method uses the acquired spatial sparse three-dimensional point cloud, and then uses Euclidean reconstruction and projective reconstruction to reconstruct the real scene of the acquired image. In 2020, Mildenhall B et al. proposed the neural radiance field NeRF, which combines deep learning and voxel rendering technology to restore high-quality three-dimensional scenes from multi-view images without explicit geometric representation. The emergence of NeRF has brought three-dimensional reconstruction to a climax, and its derivative technologies and new technical methods continue to emerge: Plenoxels in 2021 and Instant-NGP in 2022, by introducing data structures such as hash grids and sparse voxel grids, further accelerated the calculation process, and can achieve high-quality three-dimensional reconstruction without using neural networks.

[0006] 3D reconstruction methods provide strong technical support by providing accurate data support and real-time monitoring and evaluation. In recent years, they have been used in the field of forest monitoring. However, the 3D reconstruction technology involved in the field of forest monitoring still faces problems such as low resolution, small scanning range, and susceptibility to external lighting in practical applications. Among them, the high cost of NeRF in training and rendering time restricts its widespread application. Plenoxels and Instant-NGP may still face challenges when dealing with certain specific scenes or capture types. Summary of the invention

[0007] The purpose of the present invention is to address the deficiencies of existing forest monitoring technology and propose a three-dimensional forest reconstruction method based on 3DGS to monitor the forest.

[0008] Applying 3DGS technology to three-dimensional forest reconstruction can provide richer forest structure and biodiversity information, and realize accurate monitoring of forest three-dimensional structural information. The three-dimensional forest reconstruction method based on 3DGS mainly generates high-quality images through a threshold-based lateral overlap rate interception video stream image algorithm, and extracts and matches features of the image through COLMAP, sparsely reconstructs the scene, and obtains point cloud and camera pose data sets; uses 3DGS technology to train the data set, establishes a 3D Gaussian expression of the point cloud, and projects the three-dimensional Gaussian body onto a two-dimensional plane based on differentiable rasterization. The 3D Gaussian function parameters are optimized through back propagation, and the adaptive density control method is used to adapt to the geometry of different shapes in the scene. Through multiple iterative training, a high-resolution rendered forest image is obtained; finally, a visual interactive interface is presented to realize accurate monitoring of forest three-dimensional structural information.

[0009] A three-dimensional forest reconstruction method based on 3DGS technology includes the following steps: Step 1, obtaining a picture through a threshold-based lateral overlap rate video stream image capture algorithm; Step 2: Image feature extraction, matching and sparse reconstruction to obtain a data set including sparse point cloud and camera pose; Step 3: Introduce the reconstruction of forest three-dimensional model based on 3DGS technology; Step 4: Model visualization rendering to display interactive 3D scene.

[0010] Furthermore, step 1 is implemented as follows: To create a forest dataset, drones were used to shoot videos and analyze and calculate the route-related parameters. Among the built-in camera parameters, DJI air2 was used to set the camera focal length. Pixel size and resolution; in actual flight, the altitude is determined by the scene being photographed and the environment in which it is located; the stitching of UAV orthophotos requires sufficient image overlap. Since the UAV is flying around the forest and is parallel to the ground during movement, only the lateral overlap is considered. ; Use low-speed flight mode and determine the flight speed to be ; Ground sampling distance GSD, heading baseline length and flight interval length The calculation formula is as follows: ; ; in, Indicates the number of pixels parallel to the flight direction, Indicates the number of pixels perpendicular to the flight direction, the focal length of the camera , pixel size ,hanggao Substitute into formula (1) to get the ground sampling distance ; Set the ground sampling distance and set , , , Substituting into formula (2) we can obtain , ;in is 0; The captured video is applied to the threshold-based lateral overlap rate video stream image capture algorithm: the video is read through the image stream format conversion, stored in the video array and traversed, the overlap rate of adjacent images is compared, and the overlap rate threshold of the required image and the adjacent image is set to , compare the overlap rate of other pictures with the first picture, save the first picture found within the overlap rate threshold, and use the found picture as the comparison object, loop through all images, and finally get 201 two-dimensional images.

[0011] Furthermore, step 2 is specifically implemented as follows: The computer vision software COLMAP was used to extract features from a series of two-dimensional images of the forest taken by a drone. First, a new project was created in COLMAP and the folder path for storing the two-dimensional images was selected and saved. Feature extraction was selected to extract image features. When selecting the camera model, SIMPLE_PINHOLE was selected to correspond to a uniform focal length because the input image was not distorted. Feature extraction was completed for 201 two-dimensional images. Each two-dimensional image included the name, dimension, camera model, focal length, and the number of feature points in the image. Feature matching was selected for feature matching. Three-dimensional sparse reconstruction was performed on the two-dimensional images that had undergone feature extraction and feature matching. After reconstructing the scene by selecting Start reconstruction in COLMAP, the model was obtained and exported to obtain a data set, including sparse point clouds and camera poses.

[0012] Furthermore, step 3 is implemented as follows: Using 3DGS technology, based on the existing point cloud model, a learnable 3D Gaussian expression is established with each point as the center, and the splash method is used for rendering to achieve real-time high-resolution rendering of the forest 3D reconstruction, as follows: 3-1. Differentiable 3D Gaussian splash; Use anisotropic 3D Gaussian distribution as a high-quality, unstructured radiation field expression; initialize each point as a Gaussian ellipsoid, each Gaussian ellipsoid contains four types of parameter information: position information , covariance matrix, opacity and spherical harmonics; where the position information for The mean of the Gaussian; the covariance matrix is ​​used to determine the shape and direction of the Gaussian ellipsoid; opacity Represents the opacity of the Gaussian ellipsoid, which is used to render splashes; spherical harmonics represent the color of the Gaussian ellipsoid; the process of splashing is to use three-dimensional points in computer graphics for rendering and project the Gaussian ellipsoid into a plane image; 3-2. Fast differentiable rasterization; First, the entire image to be rendered is divided into blocks, and for each block, select a cone with a built-in confidence greater than The Gaussians are instantiated into Gaussian objects, which contain the numbers of the blocks they are in and the depths of the corresponding fields of view. Then the Gaussian object functions are sorted, and splashes are made to the corresponding blocks from near to far in depth order, and the accumulated traces are stacked until the opacity is saturated. , each block opens a thread block separately and runs in parallel. When the target saturation is reached in a pixel, the corresponding thread stops; finally, the thread status is queried at regular time intervals. When all pixels are saturated, the processing is terminated. After rendering the image, when the error is back-propagated, the Gaussian that needs to be optimized is found according to the above sorting; 3-3. Adaptive density control; After obtaining the rendered image, the error between the rendered image and the real image and the gradient of the corresponding parameters are calculated through the loss function. The stored Gaussian parameters are optimized and updated during back propagation using the stochastic gradient descent method, and the point cloud distribution is adaptively adjusted along the gradient direction. The "under-reconstruction" and "over-reconstruction" of the Gaussian distribution are optimized by cloning and splitting Gaussians, respectively, to better control the total number of Gaussians, and optimize the function to adapt to capturing different geometric shapes in the forest scene, thereby generating a compact representation of the scene. The Gaussian body is continuously updated with adaptive density control, and the best rendering quality rendering image is obtained after 30,000 iterations of training. 3-4. Back propagation; After obtaining the rendered image through differentiable rasterization, the error between the image and the real image is calculated through the loss function: ; in, represents the L1 loss value, express Loss value, Take 0.2 to assign weights to the two loss functions, use the stochastic gradient descent method, repeat steps 3-3, and finally reconstruct the rendered image with the best rendering quality.

[0013] Furthermore, step 4 is specifically implemented as follows: The produced forest dataset is pre-trained and rendered in real time using 3DGS technology. During the real-time rendering process, the trackball-style navigator can be selected in the floating menu on the interface. The "Scaling Modifier" option can control the size of the displayed Gaussian volume or the initial point cloud. Continuously sliding this button can present a dynamic effect and display point cloud models with different coefficients. The forest 3D model is visualized on the network viewer SIBR_remoteGaussian. When training the model, open the network viewer to visualize the scene currently being trained. An interactive 3D scene model is displayed in the interface, demonstrating the entire reconstruction process of the 3D model. The scene can be zoomed, rotated and the viewing angle can be adjusted by manipulating the mouse. The object can be moved up, down, left and right by operating the WSAD keys on the keyboard, and the scene can be freely manipulated to observe details.

[0014] The beneficial effects of the present invention are as follows: In order to obtain richer forest structure and biodiversity information and realize accurate monitoring of forest stereoscopic structure information, the present invention proposes a three-dimensional forest reconstruction method based on 3DGS. The 3DGS technology is applied to three-dimensional forest reconstruction, and the trees, terrain and other complex structures in the forest are accurately represented by Gaussian distribution, which effectively reduces reconstruction errors, improves rendering quality, and obtains more real-time and higher-quality three-dimensional reconstruction results; a threshold-based lateral overlap rate video stream image interception algorithm is proposed, and the video captured by the drone is gradually focused on the features by the algorithm to generate high-quality images by combining key information such as image overlap rate and navigation speed. By setting the overlap rate threshold of adjacent images, it can flexibly screen out pictures with appropriate similarity and adaptive adjustment according to the specific needs and conditions of the organisms and terrain in the forest video to effectively reduce redundant images and fully present the changes in forest scenes; the presentation of the visual interface realizes human-computer interaction, which is more experiential and interactive. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a complete flow chart of the forest 3D reconstruction task implemented by the present invention.

[0016] Figure 2 This is a schematic diagram of the three-dimensional reconstruction of a forest based on the 3DGS technology of the present invention.

[0017] Figure 3 A flow chart is made for the dataset used for reconstruction in the present invention. DETAILED DESCRIPTION

[0018] In order to deepen the understanding of the present invention, the method of the present invention and its detailed parameters are further specifically described below through examples.

[0019] like Figure 1-3As shown, a 3D forest reconstruction method based on 3DGS, the specific steps are as follows: Step 1: Image acquisition; To create the forest dataset, drones were used to shoot videos and the route-related parameters were analyzed and calculated. Among the built-in camera parameters, the focal length of the DJI air2 camera was used. , pixel size , the resolution is ; In actual flight, the altitude is determined by the scene being photographed and the environment in which it is located. Select a suitable viewing angle and try to collect points to finally determine the altitude. .

[0020] In order to prevent frequent acceleration and deceleration from causing a sharp drop in power, a low-speed flight mode is adopted, and the flight speed is determined to be The stitching of UAV orthophotos requires sufficient image overlap. Since the UAV flies around the forest and is parallel to the ground during movement, only the lateral overlap is considered. Since the UAV moves in a circumferential manner to acquire images, it can be considered that the obtained video is a time series of images about the plane of the outer ring of the forest, so there is a certain overlap between the frames in the video. At the same time, this overlap is related to the speed of the UAV. The faster the speed of the UAV, the lower the overlap. If the overlap is too low, the restoration effect will be poor, and if it is too high, it will cause too much data redundancy. Based on the balance between the accuracy requirements of the 3D reconstruction model and the amount of calculation, the threshold is set with a step size of 2%, and a cyclic attempt is made from 45% to 65%. The final threshold of the overlap is set to .

[0021] Ground Sampling Distance (GSD) and Heading Baseline Length and flight interval length , the calculation formula is as follows: ;

[0022] ;

[0023] in, Indicates the number of pixels parallel to the flight direction, Represents the number of pixels perpendicular to the flight direction. , , Substituting into formula (2) we get / pixel, , , , , substituting it into formula (3) to obtain , .

[0024] The 1'12'' video obtained by shooting is applied to the threshold-based lateral overlap rate video stream image capture algorithm. The video is read through the image stream format conversion (.mp4 to .jpg), stored in the video array and traversed, the overlap rate between images is compared, and the required adjacent image overlap rate threshold is set to , start to get the required pictures: first select the first picture and save it as the initial picture, compare the overlap rate of other pictures with it in time frame order, save the first picture found within the threshold, and use the newly saved picture as the comparison object, traverse the remaining pictures after the time frame, and repeat this cycle to traverse all images. Finally, 201 images are obtained, and adjacent features are gradually focused during the traversal to prepare for subsequent feature extraction and matching.

[0025] Step 2: Image feature extraction, matching and sparse reconstruction; The computer vision software COLMAP is used to extract features from a series of two-dimensional images of the forest taken by a drone. First, create a new project in COLMAP and select the folder path to store the images and save them. Select Feature extraction to extract image features. When selecting the camera model, since the input image is not distorted, select SIMPLE_PINHOLE to correspond to a uniform focal length. Complete the feature extraction of 201 images. Each image includes the name, dimension, camera model, focal length, and the number of feature points in the image. Select Feature matching for feature matching. Perform three-dimensional sparse reconstruction on the images that have been feature extracted and matched. Select Start reconstruction in COLMAP to reconstruct the scene, obtain the model and export it, and obtain a data set, including sparse point clouds and camera poses.

[0026] Step 3: Introduce the reconstruction of forest three-dimensional model based on 3DGS technology; Using 3DGS technology, based on the existing sparse point cloud model of the forest, a learnable three-dimensional Gaussian expression is established with each point as the center, and rendering is performed using the splash method to achieve real-time, high-resolution rendering of the forest three-dimensional reconstruction.

[0027] 3-1. Differentiable 3D Gaussian splash; Anisotropic 3D Gaussian distribution is used as a high-quality, unstructured radiation field expression. Each point in the forest point cloud is initialized as a Gaussian ellipsoid. Each Gaussian ellipsoid contains four types of parameter information: location information ,Right now Gaussian mean; covariance matrix, which determines the shape and direction of the Gaussian ellipsoid to match the shape and texture of the forest; opacity , represents the opacity of the Gaussian ellipsoid, which is used to render splashes; spherical harmonics, represents the color of the Gaussian ellipsoid, so that the color of the rendered forest matches the real one. The splash process, that is, rendering using three-dimensional points in computer graphics, projects the Gaussian ellipsoid into a plane image, and the predicted image of the forest can be obtained.

[0028] 3-2. Fast differentiable rasterization; In order to achieve efficient projection of the forest 3D model on a 2D plane while maintaining high quality, the 3DGS fast differentiable rasterization method is used. First, the entire forest prediction image to be rendered is divided into blocks, and for each block, select a cone with a built-in confidence greater than Gaussians are instantiated into Gaussian objects, including the numbers of the blocks they are in and the depths of the corresponding fields of view. The Gaussian object functions are sorted, and splashes are made to the corresponding blocks from near to far in depth order, and the accumulated traces are stacked until the opacity is saturated. ,Each block opens a thread block separately and runs in parallel. When the target saturation is reached in a pixel, the corresponding thread stops. The thread status is queried at regular time intervals. When all pixels are saturated, the processing is terminated. After rendering, the forest prediction image is compared with the real forest image to find the error back propagation.

[0029] 3-3. Adaptive density control; After obtaining the rendered forest prediction image, the error between it and the real forest image and the gradient of the corresponding parameters are calculated through the loss function. The stored Gaussian parameters are optimized and updated during back propagation using the stochastic gradient descent method. The initial forest point cloud distribution is adaptively adjusted along the gradient direction. In order to better fit and express the shape and texture of the forest scene, the adaptive density control method is used to handle various situations. Among them, the Gaussian distribution "under-reconstruction" (part of the forest terrain is missing) and "over-reconstruction" (part of the forest terrain diffuses to non-corresponding places) are optimized by cloning and splitting Gaussians respectively, so as to better control the total number of Gaussians and optimize the function to adapt to capturing different geometric shapes in the forest scene, thereby generating a compact representation of the forest scene. The Gaussian body is continuously updated with adaptive density control, and the forest rendering image with the best rendering quality is obtained after 30,000 iterative training.

[0030] 3-4. Back propagation; After obtaining the rendered forest prediction image through differentiable rasterization, the error between the image and the real forest image is calculated through the loss function: ;

[0031] Here Take 0.2 to assign weights to the two loss functions, use the stochastic gradient descent method, repeat steps 3-3, where a large gradient indicates a large error, and then repeat the cycle, and finally reconstruct the forest rendering image with the best rendering quality.

[0032] Step 4: Model visualization rendering; By executing the "python render.py" command in the terminal, the rendering program segment is run, and the forest dataset produced is pre-trained and rendered in real time using the 3DGS technology. Then change the working directory to SIBR_viewers in the terminal, execute the "cmake -Bbuild . -DCMAKE_BUILD_TYPE=Release" and other commands to build and manage projects across platforms and link and execute them to visualize the rendered image. During the real-time rendering process, you can choose to use the trackball-style navigator in the floating menu on the interface. The "Scaling Modifier" option can control the size of the displayed Gaussian volume or the initial point cloud. Continuously sliding this button can present a dynamic effect and display point cloud models with different coefficients. Visualize the forest 3D model on the network viewer SIBR_remoteGaussian. When training the model, open the network viewer to visualize the scene currently being trained. An interactive 3D scene model is displayed in the interface, demonstrating the entire reconstruction process of the 3D model. The scene can be zoomed, rotated and adjusted by manipulating the mouse. The object can be moved up, down, left and right by operating the WSAD keys on the keyboard, and the scene can be freely manipulated to observe details.

[0033] Specific experimental data: The 3D reconstruction of the bicycle dataset by comparing the above methods is shown in Table 1: Table 1 Comparison of various methods on bicycle dataset

[0034] As can be seen from Table 1, the 3DGS three-dimensional reconstruction method used in the present invention can train images of the same or even higher quality in a shorter time, and significantly surpasses other methods in rendering speed.

[0035] Among them, since forest images usually contain complex backgrounds and diverse textures, the threshold-based lateral overlap rate algorithm proposed in the present invention can better distinguish the foreground and background. At the same time, the algorithm can effectively reduce the storage of redundant images, with a faster processing speed and lower computing cost. The square difference matching method (TM_SQDIFF) in the template matching based on calculating the square sum of the pixel differences between the template image and the target image has a better application effect in scenes with obvious contours and local matching, but it will repeat the calculation features in the forest environment, resulting in errors in the calculation of the image overlap rate. The algorithm used in the present invention can make up for this shortcoming and is higher in computing efficiency than the algorithm; secondly, the algorithm used in the present invention has low spatial complexity, high utilization of computer resources, and can achieve fast real-time computing. If the TM_SQDIFF algorithm is used, a lot of redundant computing time will be consumed. On the contrary, the algorithm of the present invention has higher accuracy and faster speed.

[0036] Table 2 Comparison of overlapping rates of adjacent images captured by two methods

[0037] Table 2 shows the results of capturing pictures from forest videos using the threshold-based lateral overlap rate algorithm and the TM_SQDIFF algorithm. It can be seen from Table 2 that in order to make the number of pictures captured by the two algorithms consistent for comparison, the threshold is modified and a picture acquisition experiment is performed on the same forest video. The TM_SQDIFF algorithm has low discrimination and is prone to confuse different areas in the forest scene, resulting in errors in overlap rate calculation, low precision, and poor robustness to forest scenes. The threshold-based lateral overlap rate algorithm used in the present invention has clearer discrimination and higher computational efficiency.

Claims

1. A three-dimensional forest reconstruction method based on 3DGS technology, characterized in that The steps include: Step 1, obtaining an image by using a threshold-based lateral overlap rate video stream image capture algorithm; Step 2: Perform feature extraction, matching and sparse reconstruction on the image to obtain a data set including sparse point cloud and camera pose; Step 3: Construct a three-dimensional forest model based on 3DGS technology; Step 4: Model visualization rendering to display interactive 3D scene.

2. The three-dimensional forest reconstruction method based on 3DGS technology according to claim 1, characterized in that: Step 1 is implemented as follows: To create a forest dataset, drones were used to shoot videos and analyze and calculate route-related parameters; the camera focal length was set in the camera built-in parameters. , pixel size and resolution; In actual flight, the altitude Determined by the scene and environment being photographed; UAV orthophoto stitching requires sufficient image overlap. Since the drone is flying around the forest and is parallel to the ground during movement, only the lateral overlap is considered. ; Use low-speed flight mode to determine the flight speed Ground sampling distance , Heading baseline length and flight interval length The calculation formula is as follows: ； ； in, Indicates the number of pixels parallel to the flight direction, Indicates the number of pixels perpendicular to the flight direction, the focal length of the camera , pixel size ,hanggao Substitute into formula (1) to get the ground sampling distance ; Set the ground sampling distance and set , , , Substituting into formula (2) we can obtain , ;in is 0; The captured video is applied to the threshold-based lateral overlap rate video stream image capture algorithm: the video is read through the image stream format conversion, stored in the video array and traversed, the overlap rate of adjacent images is compared, and the overlap rate threshold of the required image and the adjacent image is set to , compare the overlap rates of other images with the first image, save the first image found within the overlap rate threshold, and use the found image as the comparison object, loop through all images, and finally obtain 201 two-dimensional images.

3. A three-dimensional forest reconstruction method based on 3DGS technology according to claim 1 or 2, characterized in that Step 2 is implemented as follows: The computer vision software COLMAP was used to extract features from a series of two-dimensional images of the forest taken by a drone. First, a new project was created in COLMAP and the folder path for storing the two-dimensional images was selected and saved. Feature extraction was selected to extract image features. When selecting the camera model, SIMPLE_PINHOLE was selected to correspond to a uniform focal length because the input image was not distorted. Feature extraction was completed for 201 two-dimensional images. Each two-dimensional image included the name, dimension, camera model, focal length, and the number of feature points in the image. Feature matching was selected for feature matching. Three-dimensional sparse reconstruction was performed on the two-dimensional images that had undergone feature extraction and feature matching. After selecting Start reconstruction in COLMAP to reconstruct the scene, the model was obtained and exported to obtain a data set, including sparse point clouds and camera poses.

4. The three-dimensional forest reconstruction method based on 3DGS technology according to claim 3 is characterized in that Step 3 is implemented as follows: Using 3DGS technology, based on the existing point cloud model, a learnable 3D Gaussian expression is established with each point as the center, and the splash method is used for rendering to achieve real-time high-resolution rendering of the forest 3D reconstruction, as follows: 3-1. Differentiable 3D Gaussian splash; Use anisotropic 3D Gaussian distribution as the radiation field expression; initialize each point as a Gaussian ellipsoid, each Gaussian ellipsoid contains four types of parameter information: position information , covariance matrix, opacity and spherical harmonics; where the position information for The mean of the Gaussian; the covariance matrix is ​​used to determine the shape and direction of the Gaussian ellipsoid; opacity Represents the opacity of the Gaussian ellipsoid, which is used to render splashes; spherical harmonics represent the color of the Gaussian ellipsoid; the process of splashing is to use three-dimensional points in computer graphics for rendering and project the Gaussian ellipsoid into a plane image; 3-2. Fast differentiable rasterization; First, the entire image to be rendered is divided into blocks, and for each block, select a cone with a built-in confidence greater than The Gaussians are instantiated into Gaussian objects, which contain the numbers of the blocks they are in and the depths of the corresponding fields of view. Then the Gaussian object functions are sorted, and splashes are made to the corresponding blocks from near to far in depth order, and the accumulated traces are stacked until the opacity is saturated. , each block opens a thread block separately and runs in parallel. When the target saturation is reached in a pixel, the corresponding thread stops; finally, the thread status is queried at regular time intervals. When all pixels are saturated, the processing is terminated. After rendering the image, when the error is back-propagated, the Gaussian that needs to be optimized is found according to the above sorting; 3-3. Adaptive density control; After obtaining the rendered image, the error between the rendered image and the real image and the gradient of the corresponding parameters are calculated through the loss function. The stored Gaussian parameters are optimized and updated during back propagation using the stochastic gradient descent method, and the point cloud distribution is adaptively adjusted along the gradient direction. The "under-reconstruction" and "over-reconstruction" of the Gaussian distribution are optimized by cloning and splitting Gaussians, respectively, to better control the total number of Gaussians, and optimize the function to adapt to capturing different geometric shapes in the forest scene, thereby generating a compact representation of the scene. The Gaussian body is continuously updated with adaptive density control, and the best rendering quality rendering image is obtained after 30,000 iterations of training. 3-4. Back propagation; After obtaining the rendered image through differentiable rasterization, the error between the image and the real image is calculated through the loss function: ； in, represents the L1 loss value, express Loss value, Take 0.2 to assign weights to the two loss functions, use the stochastic gradient descent method, repeat steps 3-3, and finally reconstruct the rendered image with the best rendering quality.

5. The three-dimensional forest reconstruction method based on 3DGS technology according to claim 4 is characterized in that Step 4 is implemented as follows: The produced forest dataset is pre-trained and rendered in real time using 3DGS technology. During the real-time rendering process, the trackball-style navigator can be selected in the floating menu on the interface. The "Scaling Modifier" option can control the size of the displayed Gaussian volume or the initial point cloud. Continuously sliding this button can present a dynamic effect and display point cloud models with different coefficients. The forest 3D model is visualized on the network viewer SIBR_remoteGaussian. When training the model, open the network viewer to visualize the scene currently being trained. An interactive 3D scene model is displayed in the interface, demonstrating the entire reconstruction process of the 3D model. The scene can be zoomed, rotated and the viewing angle can be adjusted by manipulating the mouse. The object can be moved up, down, left and right by operating the WSAD keys on the keyboard, and the scene can be freely manipulated to observe details.

Images