A three-dimensional scene reconstruction method, electronic device and storage medium

Abstract

The application provides a three-dimensional scene reconstruction method, an electronic device and a storage medium, and relates to the field of three-dimensional reconstruction. The method comprises the following steps: acquiring a panoramic image, panoramic camera lens parameters used for shooting the panoramic image and camera pose information; generating three-dimensional depth information by using the panoramic image, and initializing a Gaussian kernel set in a three-dimensional space according to the three-dimensional depth information; in the three-dimensional space, determining a field of view overlap region between a current frame panoramic image and a previous frame panoramic image according to the camera pose information, determining a lens overlap region by using the panoramic camera lens parameters, and determining a dynamic change region by using the current frame panoramic image and the previous frame panoramic image; updating parameters of Gaussian kernels in the Gaussian kernel set located in the field of view overlap region, the lens overlap region and the dynamic change region by using the current frame panoramic image, so as to construct a three-dimensional Gaussian splatting model; and performing three-dimensional scene reconstruction by using the three-dimensional Gaussian splatting model.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional Gaussian sputtering technology, and in particular to a three-dimensional scene reconstruction method, electronic device, and storage medium. Background Technology

[0002] 3D Gaussian Splatting (3DGS) is a 3D reconstruction technique. In related technologies, updating the Gaussian kernel of a 3D Gaussian sputtering model using an image, and then reconstructing the 3D scene based on the updated model, can easily generate a large amount of computation, severely consuming hardware resources and affecting the efficiency of real-time 3D scene reconstruction. Summary of the Invention

[0003] The purpose of this invention is to provide a three-dimensional scene reconstruction method, electronic device, and storage medium. When updating the Gaussian kernel of a three-dimensional Gaussian sputtering model, some important Gaussian kernels can be selected for updating, thereby reducing the hardware resources required for three-dimensional scene reconstruction and improving the efficiency of real-time three-dimensional scene reconstruction.

[0004] To address the aforementioned technical problems, this invention provides a three-dimensional scene reconstruction method, comprising: Acquire panoramic images, as well as the panoramic camera lens parameters and camera pose information used to capture the panoramic images; The panoramic image is used to generate three-dimensional depth information, and a Gaussian kernel set is initialized in three-dimensional space based on the three-dimensional depth information; In the three-dimensional space, the field of view overlap area between the current frame panoramic image and the previous frame panoramic image is determined according to the camera pose information, the lens overlap area is determined using the panoramic camera lens parameters, and the dynamic change area is determined using the current frame panoramic image and the previous frame panoramic image. The parameters of the Gaussian kernels located in the field of view overlap region, the lens overlap region, and the dynamic change region in the Gaussian kernel set are updated using the current frame panoramic image to construct a three-dimensional Gaussian sputtering model; The three-dimensional Gaussian sputtering model is used to reconstruct the three-dimensional scene.

[0005] Optionally, the panoramic image is captured using a panoramic camera, which includes a first lens unit and a second lens unit, and the panoramic image includes a first image captured by the first lens unit and a second image captured by the second lens unit.

[0006] Optionally, generating three-dimensional depth information using the panoramic image includes: Distortion correction, white balance calibration, and image registration are performed sequentially on the first image and the second image. The disparity map is calculated using the first image and the second image to obtain the three-dimensional depth information.

[0007] Optionally, determining the lens overlap area using the panoramic camera lens parameters includes: Obtain the first field-of-view parameters of the first lens unit and the second field-of-view parameters of the second lens unit; The lens overlap region is determined using the first field of view parameter and the second field of view parameter.

[0008] Optionally, it also includes: Pulse signals are simultaneously sent to the first lens unit and the second lens unit to synchronously acquire the first image and the second image.

[0009] Optionally, determining the field-of-view overlap region between the current frame panoramic image and the previous frame panoramic image based on the camera pose information includes: Based on the camera pose information, a first field of view cone and a second field of view cone corresponding to the previous frame panoramic image and the current frame panoramic image are determined respectively. The intersection of the first field of view cone and the second field of view cone is determined to obtain the field of view overlap region.

[0010] Optionally, determining the dynamically changing region using the current frame panoramic image and the previous frame panoramic image includes: The brightness and depth values ​​of each pixel in the current frame panoramic image and the previous frame panoramic image are obtained respectively. The brightness value is used to calculate the pixel brightness difference between the current frame panoramic image and the previous frame panoramic image, and a first target pixel whose pixel brightness difference is greater than a first preset threshold is determined. The pixel depth difference between the current frame panoramic image and the previous frame panoramic image is calculated using the depth, and a second target pixel whose pixel depth difference is greater than a second preset threshold is determined. The first target pixel and the second target pixel are integrated into a set of pixels to be processed, and noise filtering is performed on the set of pixels to be processed. Based on the panoramic camera lens parameters, each pixel in the pixel set to be processed is projected into three-dimensional space to obtain the dynamically changing region.

[0011] Optionally, it also includes: In the three-dimensional space, a three-dimensional coordinate system is set according to the optical center of the panoramic camera and the preset direction; Orthogonally divide the Gaussian kernels in the Gaussian kernel set into spatial blocks along each direction of the three-dimensional coordinate system; Determine whether the number of Gaussian kernels in the spatial block is greater than a first preset number; If the number is greater than the first preset number, the spatial block is orthogonally divided again until the number of Gaussian kernels in the spatial block is not greater than the first preset number. If it is not greater than, then determine whether the number of Gaussian kernels in the spatial block is less than the second preset number; wherein, the second preset number is less than the first preset number; If the number is less than the second preset number, the spatial block is merged with the adjacent spatial block until the number of Gaussian kernels in the spatial block is not less than the second preset number. The step of updating the parameters of the Gaussian kernels located in the field-of-view overlap region, the lens overlap region, and the dynamically changing region in the Gaussian kernel set using the current frame panoramic image includes: Determine the target spatial blocks located in the field of view overlap region, the lens overlap region, and the dynamically changing region; The Gaussian kernel parameters in each of the target spatial blocks are updated using the current frame panoramic image.

[0012] Optionally, updating the Gaussian kernel parameters in each of the target spatial blocks using the current frame panoramic image includes: Based on the projection area of ​​the target spatial block, determine the original image in the projection area in the current frame panoramic image; Based on the pixel texture information of each pixel in the original image, set the weight value of each pixel; Based on the Gaussian kernel parameters of each Gaussian kernel, the Gaussian kernels in the target space block are projected onto the projection area to obtain a projection image; Calculate the brightness difference value of each pixel between the original image and the projected image, and determine the loss value using the brightness difference value and the weight value of each pixel; The parameters of each Gaussian kernel in the target space block are updated according to the loss value. When the parameter update is completed, the Gaussian kernel parameters of the abnormal Gaussian kernels that do not meet the preset parameter constraints are adjusted using the preset parameter constraints. Determine whether the iteration exit condition is met; wherein, the iteration exit condition is that the number of iterations reaches a first preset threshold, or the decrease in the loss value between two adjacent iterations is less than a second preset threshold; If the conditions are not met, proceed to the step of projecting the Gaussian kernels in the target space block onto the projection area according to the Gaussian kernel parameters of each Gaussian kernel to obtain the projected image; If the conditions are met, the parameter update for the Gaussian kernel is completed.

[0013] Optionally, updating the Gaussian kernel parameters in each of the target spatial blocks using the current frame panoramic image includes: Multiple target spatial blocks are distributed to multiple computing units, and each computing unit uses the current frame panoramic image to perform parallel iterative updates on the Gaussian kernel in each target spatial block.

[0014] Optionally, it also includes: Gaussian kernels with transparency below a third preset threshold and three-dimensional scale less than a fourth preset threshold are filtered out, and it is determined whether the number of Gaussian kernels in the Gaussian kernel set is less than a fifth preset threshold. If it is smaller than the specified value, a Gaussian kernel is added to the overlapping field of view, the overlapping lens region, and the dynamically changing region.

[0015] The present invention also provides an electronic device, comprising: Memory, used to store computer programs; A processor is used to implement the above-described three-dimensional scene reconstruction method when executing the computer program.

[0016] The present invention also provides a computer-readable storage medium storing computer-executable instructions, which, when loaded and executed by a processor, implement the above-described three-dimensional scene reconstruction method.

[0017] This invention provides a three-dimensional scene reconstruction method, comprising: acquiring a panoramic image, and panoramic camera lens parameters and camera pose information used to capture the panoramic image; generating three-dimensional depth information using the panoramic image, and initializing a Gaussian kernel set in three-dimensional space based on the three-dimensional depth information; determining the field-of-view overlap region between the current frame panoramic image and the previous frame panoramic image in the three-dimensional space based on the camera pose information, determining the lens overlap region using the panoramic camera lens parameters, and determining the dynamically changing region using the current frame panoramic image and the previous frame panoramic image; updating the parameters of the Gaussian kernels located in the field-of-view overlap region, the lens overlap region, and the dynamically changing region in the Gaussian kernel set using the current frame panoramic image to construct a three-dimensional Gaussian sputtering model; and reconstructing the three-dimensional scene using the three-dimensional Gaussian sputtering model.

[0018] The beneficial effects of this invention are as follows: Firstly, it acquires a panoramic image, along with the panoramic camera lens parameters and camera pose information used to capture the panoramic image. Then, it generates three-dimensional depth information using the panoramic image and initializes a Gaussian kernel set in three-dimensional space based on this depth information. Before updating the Gaussian kernel parameters, this invention determines the field-of-view overlap region between the current frame panoramic image and the previous frame panoramic image in three-dimensional space based on the camera pose information, determines the lens overlap region using the panoramic camera lens parameters, and determines the dynamically changing region using the current frame panoramic image and the previous frame panoramic image. Only the Gaussian kernels located in the field-of-view overlap region, lens overlap region, and dynamically changing region within the Gaussian kernel set are updated using the current frame panoramic image to construct a three-dimensional Gaussian sputtering model. Thus, when updating the Gaussian kernels of the three-dimensional Gaussian sputtering model using the panoramic image, this invention only updates the Gaussian kernels located in the field-of-view overlap region and the significantly changing region in the current frame panoramic image, without updating all Gaussian kernels in the model. This effectively reduces the amount of parameter updates required for the three-dimensional Gaussian sputtering model, thereby improving the efficiency of real-time three-dimensional scene reconstruction.

[0019] The present invention also provides an electronic device and a computer-readable storage medium, which have the above-mentioned beneficial effects. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0021] Figure 1 A flowchart of a three-dimensional scene reconstruction method provided in an embodiment of the present invention; Figure 2 This is a structural block diagram of a three-dimensional scene reconstruction device provided in an embodiment of the present invention; Figure 3 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0023] 3D Gaussian Splatting (3DGS) is a 3D reconstruction technique. In related technologies, when updating the Gaussian kernel of a 3D Gaussian sputtering model using an image, and then reconstructing a 3D scene based on the updated model, a full update of the Gaussian kernel is typically required. This can generate significant computational overhead, severely consuming hardware resources and impacting the efficiency of real-time 3D scene reconstruction.

[0024] In view of this, to address the technical problem of how to improve the efficiency of 3D scene reconstruction, the present invention provides a 3D scene reconstruction method that can select some important Gaussian kernels for updating when updating the Gaussian kernel of a 3D Gaussian sputtering model, thereby reducing the hardware resources required for 3D scene reconstruction and improving the efficiency of real-time 3D scene reconstruction.

[0025] Please refer to Figure 1 , Figure 1 A flowchart of a three-dimensional scene reconstruction method provided in an embodiment of the present invention, the method may include: S10. Acquire panoramic images, as well as the panoramic camera lens parameters and camera pose information used to capture the panoramic images.

[0026] In this step, a panoramic image refers to image data covering the entire field of view, which can be acquired using a panoramic camera. A panoramic camera may contain multiple lens units, allowing it to acquire image data from different perspectives, which can then be stitched together to form a complete panoramic image. In one possible scenario, the panoramic camera may contain a first lens unit and a second lens unit, and the panoramic image may consist of a first image acquired by the first lens unit and a second image acquired by the second lens unit. For example, the panoramic camera may employ a dual panoramic lens group arranged symmetrically front and back, and the panoramic image may be composed of a forward panoramic image and a backward panoramic image.

[0027] Panoramic camera lens parameters can include internal parameters of the lens unit (such as focal length, principal point coordinates (u0, v0), viewing angle, exposure parameters, etc.) as well as extrinsic parameter calibration results between lens units (such as rotation matrix R, translation vector T, etc.). Based on the panoramic camera lens parameters, image data acquired by each lens unit can be calibrated and registered, and the viewing angle range of each lens unit can be determined to facilitate the subsequent selection of important Gaussian kernels to be updated.

[0028] Camera pose information refers to the position and orientation of the panoramic camera when capturing panoramic images. It is an important piece of information required for 3D scene reconstruction and can also be used to filter important Gaussian kernels to be updated.

[0029] It is worth noting that traditional panoramic cameras often use asynchronous triggering for multi-view acquisition, resulting in time delays between different lenses. This causes object positions to shift in dynamic scenes, leading to data time inconsistencies and consequently causing problems such as blurred reconstruction and misalignment. To address this, this embodiment can incorporate a hardware synchronization triggering unit within the panoramic camera. This unit can simultaneously send pulse signals to both the first and second lens units to synchronously acquire dual-view panoramic data with consistent timestamps.

[0030] Based on this, the method may also include: S11: Simultaneously send pulse signals to the first lens unit and the second lens unit to synchronously acquire the first image and the second image.

[0031] S20. Generate three-dimensional depth information using panoramic images, and initialize a Gaussian kernel set in three-dimensional space based on the three-dimensional depth information.

[0032] In this step, since the panoramic image can be composed of a first image captured by the first lens unit and a second image captured by the second lens unit, there is a physical parallax between the first lens unit and the second lens unit. Therefore, based on the parallax characteristics, a parallax map can be calculated using the first and second images to obtain the three-dimensional depth information in the three-dimensional scene captured by the panoramic camera, such as the three-dimensional depth information of the photographed object.

[0033] A Gaussian kernel (or Gaussian ellipsoid) is the basic unit for representing a 3D scene. Its core definition is a 3D ellipsoid parameterized by multidimensional attributes, used to model the geometry and appearance of a local region within the scene. A Gaussian kernel can have multiple parameters, such as position, size, rotation quaternion, opacity, and spherical harmonic coefficients. To improve the 3D scene reconstruction effect, the parameters of the Gaussian kernel need to be iteratively updated multiple times based on the acquired panoramic images. 3D depth information can be used to initialize a set of Gaussian kernels in 3D space. For example, a 3D point cloud can be generated based on the 3D depth information, and a Gaussian kernel can be generated based on each point in the 3D point cloud, resulting in a set of Gaussian kernels.

[0034] The disparity map can be calculated using the semi-global matching (SGM) algorithm, which combines the extrinsic parameter calibration results of multiple lens units (rotation matrix R, translation vector T) to output sub-pixel level disparity accuracy, providing depth constraints for the initialization of the 3D Gaussian kernel.

[0035] Of course, since the first and second lens units may have lens distortion and inconsistent white balance, the first and second images can be preprocessed before calculating the disparity map. This preprocessing includes distortion correction, white balance calibration, and image registration, performed sequentially on the first and second images. Based on this, generating 3D depth information from panoramic images can include: S211: Perform distortion correction, white balance calibration, and image registration on the first and second images sequentially; S212: Calculate the disparity map using the first and second images to obtain three-dimensional depth information.

[0036] Furthermore, to facilitate the selection of Gaussian kernels to be updated, and to facilitate the parallel updating of Gaussian kernels by multiple computing units (such as multiple GPUs), this embodiment can also divide the three-dimensional space into blocks after generating the Gaussian kernel set, so as to divide the Gaussian kernels into various spatial blocks.

[0037] Specifically, using the panoramic camera's coordinate system (with the optical center as the origin and the forward direction as the Z-axis) as a reference, the 3D space can be orthogonally divided along the X / Y / Z axes to create spatial blocks of preset sizes. The preset base size for each spatial block can be 1m × 1m × 1m, and this size can be adjusted as needed or dynamically adjusted according to the scene scale; for example, a close-up scene can be scaled down to 0.5m × 0.5m × 0.5m. Alternatively, to adapt to the spherical field of view of a panoramic camera, "spherical coordinate segmentation" can be used, i.e., dividing the space according to azimuth / elevation / distance.

[0038] Subsequently, after the initial segmentation is completed, this embodiment can continue to perform dynamic subdivision and merging: if the number of Gaussian kernels in a certain segment is greater than 5000 (threshold), then a second subdivision is performed; if it is less than 100, then it is merged with the adjacent segment, ultimately ensuring that the number of Gaussian kernels in each segment is in the range of [200, 5000], balancing parallel efficiency and computational load.

[0039] Subsequently, after ensuring that the number of Gaussian kernels in each block is within a preset range, this embodiment can assign a unique three-dimensional index to each block, such as (X_idx, Y_idx, Z_idx), and can establish a "block-ID-Gaussian kernel" mapping table to record the Gaussian kernel IDs contained in each block.

[0040] Based on this, the method may also include: S221: In three-dimensional space, set up a three-dimensional coordinate system based on the optical center of the panoramic camera and the preset direction; S222: Perform orthogonal partitioning along each direction of the three-dimensional coordinate system, dividing the Gaussian kernels in the Gaussian kernel set into each spatial partition; S223: Determine whether the number of Gaussian kernels in the spatial block is greater than the first preset number; S224: If it is greater than, then the spatial block is orthogonally divided again until the number of Gaussian kernels in the spatial block is not greater than the first preset number; S225: If not greater than, determine whether the number of Gaussian kernels in the spatial block is less than the second preset number; wherein, the second preset number is less than the first preset number; S226: If it is less than, then merge the spatial block with the adjacent spatial block until the number of Gaussian kernels in the spatial block is not less than the second preset number.

[0041] S30. In three-dimensional space, determine the field of view overlap area between the current frame panoramic image and the previous frame panoramic image based on the camera pose information, determine the lens overlap area using the panoramic camera lens parameters, and determine the dynamically changing area using the current frame panoramic image and the previous frame panoramic image.

[0042] In the construction of a 3D Gaussian sputtering model, each time a panoramic image is obtained, multiple rounds of parameter updates need to be performed on all Gaussian kernels in the model using that image. Related technologies require a full parameter update of all Gaussian kernels in the model, resulting in a large amount of data processing, potentially wasting hardware resources and impacting the efficiency of real-time 3D scene reconstruction.

[0043] Therefore, to avoid a full update of the Gaussian kernel, this step will select important Gaussian kernels and update their parameters only. Specifically, this embodiment can focus on selecting Gaussian kernels in the current frame of the panoramic image that have undergone incremental updates compared to the previous frame for parameter updates. This embodiment can determine the location of Gaussian kernels with incremental updates from the following three aspects: 1. Determine the overlap area of ​​the field of view between the current frame and the previous frame: This embodiment can read the camera pose information of the current frame panoramic image and the previous frame panoramic image (associated with timestamps and combined with offline calibrated intrinsic / extrinsic parameters), and calculate the field of view (FOV) cones of the two frames based on the camera pose information. Subsequently, the intersection of the two FOV cones can be determined to obtain the FOV overlap region in three-dimensional space (represented by an axis-aligned bounding box AABB, containing min / max (X,Y,Z)).

[0044] Based on this, determining the field-of-view overlap region between the current frame panoramic image and the previous frame panoramic image using camera pose information can include: S311: Based on the camera pose information, determine the first field of view cone and the second field of view cone corresponding to the previous panoramic image and the current panoramic image, respectively; S312: Determine the intersection of the first field of view cone and the second field of view cone to obtain the field of view overlap region.

[0045] 2. Supplement the lens overlap area of ​​multi-lens units: By combining the extrinsic parameters (R, T) of multiple lens units, the intersection of the fields of view between multiple lens units is calculated to obtain the lens overlap area.

[0046] Based on this, determining the lens overlap area using panoramic camera lens parameters can include: S321: Obtain the first field-of-view parameters of the first lens unit and the second field-of-view parameters of the second lens unit; S322: Determine the lens overlap area using the first field of view parameter and the second field of view parameter.

[0047] After obtaining the aforementioned overlapping regions of the field of view and the lens, the three-dimensional coordinates (x, y, z) of all Gaussian kernels can be traversed, and it can be determined whether the coordinates fall within the AABB range of the overlapping region. If the conditions are met, they are entered into the "Overlapping Region Gaussian Kernel List" and associated with the corresponding block.

[0048] 3. Determine the dynamic change area: In this embodiment, dynamically changing regions refer to areas in the panoramic image where pixel brightness and pixel depth exhibit significant changes. These regions commonly experience object movement and lighting changes, requiring focused updates to the relevant Gaussian kernels. The specific detection steps are as follows: 1) 2D layer dynamic mask detection: Photometric difference: Calculate the pixel brightness difference ΔI=|I_current-I_prev| between the current frame and the previous frame, set a threshold (ΔI>20 / 255, i.e. 8% brightness difference), and obtain the photometric difference mask; Depth difference: Calculate the pixel-level difference ΔD=|D_current-D_prev| of the depth prior, set a threshold (ΔD>5%×D_current), and obtain the depth difference mask; The union of the two is used to obtain a two-dimensional dynamic mask, and then noise is filtered by optical flow (regions with optical flow vector magnitude < 1 pixel are removed).

[0049] 2) 3D spatial extent mapping: The pixel coordinates of the 2D dynamic mask are projected back into 3D space using camera intrinsic parameters (focal length f, principal point (u0, v0)) to obtain the three-dimensional AABB range of the dynamic region.

[0050] 3) Dynamic region Gaussian kernel selection: Traverse the 3D coordinates of the Gaussian kernel, determine whether it falls within the 3D range of the dynamic region, and enter the "Dynamic Region Gaussian Kernel List" if it meets the conditions, and associate it with the corresponding block.

[0051] Based on this, determining dynamically changing regions using the current frame panoramic image and the previous frame panoramic image can include: S331: Obtain the brightness and depth values ​​of each pixel in the current frame panoramic image and the previous frame panoramic image respectively; S332: Calculate the pixel brightness difference between the current frame panoramic image and the previous frame panoramic image using the brightness value, and determine the first target pixel whose pixel brightness difference is greater than the first preset threshold. S333: Calculate the pixel depth difference between the current frame panoramic image and the previous frame panoramic image using depth to determine the second target pixel whose pixel depth difference is greater than the second preset threshold; S334: Integrate the first target pixel and the second target pixel into a pixel set to be processed, and perform noise filtering on the pixel set to be processed; S335: Based on the panoramic camera lens parameters, project each pixel in the pixel set to be processed onto three-dimensional space to obtain the dynamically changing area.

[0052] S40. Use the current frame panoramic image to update the parameters of the Gaussian kernels located in the field of view overlap region, lens overlap region, and dynamic change region in the Gaussian kernel set to construct a three-dimensional Gaussian sputtering model.

[0053] After obtaining the aforementioned overlapping fields of view, lens overlap areas, and dynamically changing areas, this embodiment can mark only the blocks containing the overlapping and dynamic areas as "to be optimized," while the rest are "dormant" and do not participate in the current update. Therefore, when updating Gaussian kernel parameters using the current frame panoramic image, this embodiment can significantly reduce the number of Gaussian kernels involved in the update, thereby reducing the processing load required to construct the 3D Gaussian sputtering model and improving the efficiency of real-time 3D scene reconstruction.

[0054] Based on this, the parameters of the Gaussian kernels located in the field-of-view overlap region, lens overlap region, and dynamic change region of the Gaussian kernel set are updated using the current frame panoramic image, including: S41: Determine the target spatial blocks located in the field of view overlap region, lens overlap region, and dynamic change region; S42: Update the parameters of the Gaussian kernel in each target spatial block using the panoramic image of the current frame.

[0055] The following describes the specific process of updating the Gaussian kernel parameters: 1. Objective function construction: For each block, the following objective function can be constructed: ; Where E represents the objective function. The spatial blocks are mapped onto the two-dimensional plane containing the current frame's panoramic image to obtain the projection region, where P represents the set of two-dimensional pixels within this projection region, and w... p This represents the pixel weight, which can be set based on the texture information in the current frame's panoramic image. The weight for texture-rich areas is ≥0.8, and the weight for texture-sparse areas is ≤0.3. Image texture can be calculated based on pixel gradient variance. prender I represents the rendering brightness of the Gaussian kernel projection. preal This represents the true brightness of the original image.

[0056] 2. Incremental parameter update: Optimizer initialization: Momentum gradient descent (SGD) is used with a momentum coefficient of 0.9 and an initial learning rate of 1e-4 (decreasing by 10% per iteration). Only update the core parameters: position (x,y,z), scale (σ1,σ2,σ3), rotation (quaternion q), and transparency (α); Block iteration: Each block is iterated independently for ≤10 times. If the photometric error decreases by <1%, the iteration is terminated early. Parameter constraints: After the update, the scale must be ≥0, the transparency α∈[0,1], and the rotation quaternion must be normalized to ensure physical validity.

[0057] 3. Global parameter synchronization: After all the blocks to be optimized have been iterated, the global Gaussian kernel parameter library is updated synchronously, while the Gaussian kernel parameters of the dormant blocks remain unchanged.

[0058] Based on this, updating the Gaussian kernel parameters in each target spatial block using the current frame panoramic image can include: S421: Determine the original image in the projection area of ​​the current frame panoramic image based on the projection area of ​​the target space block; S422: Set the weight value of each pixel based on the pixel texture information of each pixel in the original image; S423: Based on the Gaussian kernel parameters of each Gaussian kernel, project the Gaussian kernels in the target space block to the projection area to obtain the projected image; S424: Calculate the brightness difference value of each pixel between the original image and the projected image, and determine the loss value using the brightness difference value and the weight value of each pixel; S425: Update the parameters of each Gaussian kernel in the target space block according to the loss value, and when the parameter update is completed, adjust the Gaussian kernel parameters of abnormal Gaussian kernels that do not meet the preset parameter constraints using the preset parameter constraints. S426: Determine whether the iteration exit condition is met; wherein, the iteration exit condition is that the number of iterations reaches the first preset threshold, or the decrease in the loss value between two adjacent iterations is less than the second preset threshold; S427: If not satisfied, proceed to the step of projecting the Gaussian kernels in the target space block to the projection area according to the Gaussian kernel parameters of each Gaussian kernel to obtain the projected image. S428: If satisfied, the parameter update of the Gaussian kernel is completed.

[0059] It is worth noting that since this embodiment has divided the three-dimensional space into blocks and can divide the Gaussian kernel into each spatial block, this embodiment can send multiple target spatial blocks to multiple computing units. Each computing unit can use the current frame panoramic image to perform parallel iterative updates on the Gaussian kernel in each target spatial block, thereby achieving parallel processing and improving processing efficiency.

[0060] Specifically, based on the "block-Gaussian kernel" mapping table, blocks containing Gaussian kernels to be updated (blocks to be optimized) can be selected; then, an independent GPU thread can be allocated to each block to be optimized to ensure that the update time per frame remains within a controllable range.

[0061] S50. Reconstructing a 3D scene using a 3D Gaussian sputtering model.

[0062] In this embodiment, an optimized 3D scene Gaussian model can be projected onto the image plane using a GPU-accelerated Gaussian sputtering pipeline to generate a full-view 3D reconstruction result, which is then output to a display device or storage module in real time. Additionally, when hardware computing power is insufficient, the number of Gaussian cores is automatically reduced (e.g., a minimum of ≥500,000) to maintain the rendering frame rate.

[0063] Based on the above embodiments, the present invention first acquires a panoramic image, as well as the panoramic camera lens parameters and camera pose information used to capture the panoramic image. Then, it generates three-dimensional depth information using the panoramic image and initializes a Gaussian kernel set in three-dimensional space based on the three-dimensional depth information. Before updating the Gaussian kernel parameters, the present invention determines the field-of-view overlap region between the current frame panoramic image and the previous frame panoramic image in three-dimensional space based on the camera pose information, determines the lens overlap region using the panoramic camera lens parameters, and determines the dynamically changing region using the current frame panoramic image and the previous frame panoramic image. It then updates the parameters of Gaussian kernels located in the field-of-view overlap region, lens overlap region, and dynamically changing region using only the current frame panoramic image, thereby constructing a three-dimensional Gaussian sputtering model. Thus, when updating the three-dimensional Gaussian sputtering model using the input image, the present invention can update the parameters of Gaussian kernels only located in the field-of-view overlap region and the significantly changing region in the current frame panoramic image, without updating all Gaussian kernels in the model. Therefore, only important Gaussian kernels can be updated, effectively reducing the number of parameters required to update the three-dimensional Gaussian sputtering model, which is beneficial for real-time three-dimensional reconstruction.

[0064] Based on the above embodiments, in order to avoid Gaussian kernels with low transparency and small three-dimensional scale interfering with parameter updates and the efficiency of three-dimensional scene reconstruction, this embodiment can also filter out Gaussian kernels with low transparency and small three-dimensional scale in real time, and can replenish new Gaussian kernels in a timely manner to maintain the stability of the total number of Gaussian kernels.

[0065] Based on this, the method may also include: S61: Filter out Gaussian kernels with transparency lower than the third preset threshold and three-dimensional scale smaller than the fourth preset threshold, and determine whether the number of Gaussian kernels in the Gaussian kernel set is less than the fifth preset threshold.

[0066] Specifically, this embodiment can remove Gaussian kernels with transparency below the threshold (0.01) or scale smaller than the pixel-level threshold (1 pixel corresponds to a three-dimensional spatial scale) in real time to ensure real-time performance.

[0067] S62: If it is less than 1, a Gaussian kernel will be added to the field of view overlap region, lens overlap region, and dynamic change region.

[0068] After a single frame is cropped, if the total number of Gaussian kernels is lower than the lower edge of the preset range (e.g., the normal range is 1 million to 5 million, with the lower edge being 1 million; the lower limit is 500,000 when hardware computing power is insufficient), or if there are valid areas in the scene that are not covered by Gaussian kernels, on-demand addition will be triggered immediately.

[0069] This embodiment can add Gaussian kernels only in the overlapping areas of the current frame and the previous frame, in areas of dynamic scene changes, and in areas with rich textures, instead of adding them to the entire scene, thereby ensuring real-time performance.

[0070] In addition, the newly added Gaussian kernel can be generated in the camera coordinate system based on the depth prior information (disparity map transformation) output by the data preprocessing module, and associated with core parameters such as position (x,y,z), scale (σ1,σ2,σ3), rotation quaternion (q), and transparency (α), ensuring the compatibility of the new kernel with the original Gaussian model without the need for additional re-initialization.

[0071] In addition, the number of Gaussian kernels added in a single frame is less than or equal to the number of kernels removed in a single frame, and the total number of kernels added globally does not exceed the upper edge of a preset range (e.g., 5 million), which allows for precise control of "how much is lost and how much is added".

[0072] Based on the above embodiments, the three-dimensional scene reconstruction method provided by the present invention will be fully described below: S1: Synchronous Acquisition: Through the hardware synchronization trigger unit of the panoramic synchronous acquisition module, the front and rear dual panoramic lens groups are controlled to acquire the front panoramic image and the rear panoramic image of the scene at the same time, and output the dual-view panoramic raw data with timestamps. S2: Data Preprocessing: Perform the following operations on the raw dual-view panoramic data: S21: The Zhang calibration method is used for distortion correction to eliminate radial and tangential distortion of the fisheye lens; S22: Perform white balance calibration and brightness normalization to ensure consistent brightness in dual-view images; S23: Perform image registration based on pre-calibrated extrinsic parameters (R, T) to align the pixel coordinate system of the dual-view images; S24: Calculate the disparity map using the SGM algorithm and convert it into depth prior information; S3: 3DGS Real-time Reconstruction: S31: Gaussian initialization: Based on the depth prior information in step S24, an initial 3D Gaussian kernel set is generated in the camera coordinate system. Each Gaussian kernel is associated with position (x,y,z), scale (σ1,σ2,σ3), rotation (quaternion q), and transparency (α) parameters. S32: Incremental parameter update: Input standardized multi-view image data into the 3DGS model, with the goal of minimizing multi-view photometric error, and use a block optimization strategy to update the Gaussian parameters, only optimizing the Gaussian kernel in the overlapping and dynamic regions, with the number of iterations ≤10. S33: Gaussian kernel clipping: Removes invalid Gaussian kernels (transparency < 0.01 or scale < 1 pixel), maintaining the total number of Gaussian kernels within a preset range; S4: Real-time rendering: Through the GPU-accelerated Gaussian sputtering pipeline, the optimized 3D scene Gaussian model is projected onto the image plane to generate a full-view 3D reconstruction result, which is then output to the display device or storage module in real time. S5: Iterative Loop: Repeat steps S1 to S4 to achieve continuous real-time 3D reconstruction of dynamic scenes.

[0073] The following describes the three-dimensional scene reconstruction device, electronic device, computer program product, and computer-readable storage medium provided in the embodiments of the present invention. The three-dimensional scene reconstruction device, electronic device, computer program product, and computer-readable storage medium described below can be referred to in correspondence with the three-dimensional scene reconstruction method described above.

[0074] Please refer to Figure 2 , Figure 2 This is a structural block diagram of a three-dimensional scene reconstruction device provided in an embodiment of the present invention. The device may include: The acquisition module 201 is used to acquire panoramic images, as well as the panoramic camera lens parameters and camera pose information used to capture the panoramic images; The Gaussian kernel initialization module 202 is used to generate three-dimensional depth information using panoramic images and initialize a set of Gaussian kernels in three-dimensional space based on the three-dimensional depth information. The incremental region determination module 203 is used to determine the field of view overlap region between the current frame panoramic image and the previous frame panoramic image in three-dimensional space based on camera pose information, determine the lens overlap region using panoramic camera lens parameters, and determine the dynamically changing region using the current frame panoramic image and the previous frame panoramic image. The parameter update module 204 is used to update the parameters of the Gaussian kernels in the Gaussian kernel set located in the field of view overlap region, lens overlap region, and dynamic change region using the current frame panoramic image, so as to construct a three-dimensional Gaussian sputtering model. The 3D reconstruction module 205 is used to reconstruct 3D scenes using a 3D Gaussian sputtering model.

[0075] Optionally, the panoramic image is captured using a panoramic camera, which includes a first lens unit and a second lens unit, and the panoramic image includes a first image captured by the first lens unit and a second image captured by the second lens unit.

[0076] Optionally, the Gaussian kernel initialization module 202 includes: The preprocessing submodule is used to perform distortion correction, white balance calibration, and image registration on the first image and the second image in sequence. The disparity map calculation submodule is used to calculate the disparity map using the first image and the second image to obtain three-dimensional depth information.

[0077] Optionally, the incremental region determination module 203 includes: The lens overlap region determination submodule is used to: obtain the first field-of-view parameters of the first lens unit and the second field-of-view parameters of the second lens unit; and determine the lens overlap region using the first field-of-view parameters and the second field-of-view parameters.

[0078] Optionally, the device further includes: The image acquisition module is used to simultaneously send pulse signals to the first lens unit and the second lens unit to synchronously acquire the first image and the second image.

[0079] Optionally, the incremental region determination module 203 includes: The field-of-view overlap region determination submodule is used to determine the first field-of-view cone and the second field-of-view cone corresponding to the previous frame panoramic image and the current frame panoramic image, respectively, based on the camera pose information; and to determine the intersection of the first field-of-view cone and the second field-of-view cone to obtain the field-of-view overlap region.

[0080] Optionally, the incremental region determination module 203 includes: The dynamic change region determination submodule is used to obtain the brightness and depth values ​​of each pixel in the current frame panoramic image and the previous frame panoramic image, respectively. The pixel brightness difference between the current frame panoramic image and the previous frame panoramic image is calculated using the brightness value, and the first target pixel whose pixel brightness difference is greater than the first preset threshold is determined. The depth difference between the current frame panoramic image and the previous frame panoramic image is calculated using depth, and a second target pixel with a pixel depth difference greater than a second preset threshold is determined. The first target pixel and the second target pixel are integrated into a pixel set to be processed, and noise filtering is performed on the pixel set to be processed. Based on the panoramic camera lens parameters, each pixel in the pixel set to be processed is projected into three-dimensional space to obtain the dynamically changing region.

[0081] Optionally, the device may further include: Spatial partitioning module, used for: In three-dimensional space, a three-dimensional coordinate system is set according to the optical center of the panoramic camera and the preset direction; Orthogonally divide the Gaussian kernels in the Gaussian kernel set into spatial blocks along each direction of the three-dimensional coordinate system; Determine whether the number of Gaussian kernels in the spatial block is greater than a first preset number; If the number is greater than the first preset number, the spatial block is orthogonally divided again until the number of Gaussian kernels in the spatial block is not greater than the first preset number. If it is not greater than, then determine whether the number of Gaussian kernels in the spatial block is less than the second preset number; wherein, the second preset number is less than the first preset number; If the number is less than the second preset number, the spatial block is merged with the adjacent spatial block until the number of Gaussian kernels in the spatial block is not less than the second preset number. Parameter update module 204 can be used for: Identify the target spatial blocks located in the overlapping regions of the field of view, the overlapping regions of the lenses, and the dynamically changing regions; The parameters of the Gaussian kernel in each target spatial block are updated using the panoramic image of the current frame.

[0082] Optionally, the parameter update module 204 can be used for: Based on the projection area of ​​the target spatial block, determine the original image in the projection area in the current frame panoramic image; Set the weight value of each pixel based on the pixel texture information of each pixel in the original image; Based on the Gaussian kernel parameters of each Gaussian kernel, the Gaussian kernels in the target space block are projected onto the projection area to obtain the projected image; Calculate the brightness difference value of each pixel between the original image and the projected image, and determine the loss value using the brightness difference value and the weight value of each pixel; The parameters of each Gaussian kernel in the target space block are updated according to the loss value. When the parameter update is completed, the Gaussian kernel parameters of the abnormal Gaussian kernels that do not meet the preset parameter constraints are adjusted using the preset parameter constraints. Determine whether the iteration exit condition is met; wherein, the iteration exit condition is that the number of iterations reaches a first preset threshold, or the decrease in the loss value between two adjacent iterations is less than a second preset threshold; If the conditions are not met, proceed to the step of projecting the Gaussian kernels in the target space block to the projection area according to the Gaussian kernel parameters of each Gaussian kernel to obtain the projected image. If the conditions are met, the parameter update for the Gaussian kernel is completed.

[0083] Optionally, the parameter update module 204 can be used for: Multiple target spatial blocks are distributed to multiple computing units, and each computing unit uses the current frame panoramic image to perform parallel iterative updates of the Gaussian kernel in each target spatial block.

[0084] Optionally, the device may further include: The Gaussian kernel screening and supplementation module is used for: Gaussian kernels with transparency below the third preset threshold and three-dimensional scale below the fourth preset threshold are filtered out, and it is determined whether the number of Gaussian kernels in the Gaussian kernel set is less than the fifth preset threshold. If it is smaller than 0, a new Gaussian kernel will be added to the overlapping areas of the field of view, the overlapping areas of the lenses, and the areas of dynamic change.

[0085] Please refer to Figure 3 , Figure 3 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. The present invention provides an electronic device 10, including a processor 11 and a memory 12; wherein, the memory 12 is used to store a computer program; the processor 11 is used to execute the three-dimensional scene reconstruction method provided in the foregoing embodiment when executing the computer program.

[0086] For details regarding the specific process of the above-mentioned 3D scene reconstruction method, please refer to the relevant content provided in the foregoing embodiments, which will not be repeated here.

[0087] Furthermore, the memory 12, as a carrier for resource storage, can be a read-only memory, random access memory, disk, or optical disk, and the storage method can be temporary storage or permanent storage.

[0088] In addition, the electronic device 10 also includes a power supply 13, a communication interface 14, an input / output interface 15, and a communication bus 16; wherein, the power supply 13 is used to provide operating voltage for each hardware device on the electronic device 10; the communication interface 14 can create a data transmission channel between the electronic device 10 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this invention, and is not specifically limited here; the input / output interface 15 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0089] This invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the three-dimensional scene reconstruction method described in the above embodiments.

[0090] Since the embodiments of the computer program product section correspond to the embodiments of the 3D scene reconstruction method section, please refer to the description of the embodiments of the 3D scene reconstruction method section for the embodiments of the computer program product section, and they will not be repeated here.

[0091] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the three-dimensional scene reconstruction method described in the above embodiments.

[0092] Since the embodiments of the computer-readable storage medium portion correspond to the embodiments of the three-dimensional scene reconstruction method portion, the embodiments of the storage medium portion are described in the description of the embodiments of the three-dimensional scene reconstruction method portion, and will not be repeated here.

[0093] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0094] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0095] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0096] The present invention has provided a detailed description of a three-dimensional scene reconstruction method, electronic device, and storage medium. Specific examples have been used to illustrate the principles and implementation methods of the invention. The descriptions of these embodiments are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to the invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the present invention.

Claims

1. A method for reconstructing a three-dimensional scene, characterized in that, include: Acquire panoramic images, as well as the panoramic camera lens parameters and camera pose information used to capture the panoramic images; The panoramic image is used to generate three-dimensional depth information, and a Gaussian kernel set is initialized in three-dimensional space based on the three-dimensional depth information; In the three-dimensional space, the field of view overlap area between the current frame panoramic image and the previous frame panoramic image is determined according to the camera pose information, the lens overlap area is determined using the panoramic camera lens parameters, and the dynamic change area is determined using the current frame panoramic image and the previous frame panoramic image. The parameters of the Gaussian kernels located in the field of view overlap region, the lens overlap region, and the dynamic change region in the Gaussian kernel set are updated using the current frame panoramic image to construct a three-dimensional Gaussian sputtering model; The three-dimensional Gaussian sputtering model is used to reconstruct the three-dimensional scene.

2. The three-dimensional scene reconstruction method according to claim 1, characterized in that, The panoramic image is captured using a panoramic camera, which includes a first lens unit and a second lens unit. The panoramic image includes a first image captured by the first lens unit and a second image captured by the second lens unit.

3. The three-dimensional scene reconstruction method according to claim 2, characterized in that, The process of generating three-dimensional depth information using the panoramic image includes: Distortion correction, white balance calibration, and image registration are performed sequentially on the first image and the second image. The disparity map is calculated using the first image and the second image to obtain the three-dimensional depth information.

4. The three-dimensional scene reconstruction method according to claim 2, characterized in that, The step of determining the lens overlap area using the panoramic camera lens parameters includes: Obtain the first field-of-view parameters of the first lens unit and the second field-of-view parameters of the second lens unit; The lens overlap region is determined using the first field of view parameter and the second field of view parameter.

5. The three-dimensional scene reconstruction method according to claim 2, characterized in that, Also includes: Pulse signals are simultaneously sent to the first lens unit and the second lens unit to synchronously acquire the first image and the second image.

6. The three-dimensional scene reconstruction method according to claim 1, characterized in that, Determining the field-of-view overlap region between the current frame panoramic image and the previous frame panoramic image based on the camera pose information includes: Based on the camera pose information, a first field of view cone and a second field of view cone corresponding to the previous frame panoramic image and the current frame panoramic image are determined respectively. The intersection of the first field of view cone and the second field of view cone is determined to obtain the field of view overlap region.

7. The three-dimensional scene reconstruction method according to claim 1, characterized in that, The step of determining dynamically changing regions using the current frame panoramic image and the previous frame panoramic image includes: The brightness and depth values ​​of each pixel in the current frame panoramic image and the previous frame panoramic image are obtained respectively. The brightness value is used to calculate the pixel brightness difference between the current frame panoramic image and the previous frame panoramic image, and a first target pixel whose pixel brightness difference is greater than a first preset threshold is determined. The pixel depth difference between the current frame panoramic image and the previous frame panoramic image is calculated using the depth, and a second target pixel whose pixel depth difference is greater than a second preset threshold is determined. The first target pixel and the second target pixel are integrated into a set of pixels to be processed, and noise filtering is performed on the set of pixels to be processed. Based on the panoramic camera lens parameters, each pixel in the pixel set to be processed is projected into three-dimensional space to obtain the dynamically changing region.

8. The three-dimensional scene reconstruction method according to claim 1, characterized in that, Also includes: In the three-dimensional space, a three-dimensional coordinate system is set according to the optical center of the panoramic camera and the preset direction; Orthogonally divide the Gaussian kernels in the Gaussian kernel set into spatial blocks along each direction of the three-dimensional coordinate system; Determine whether the number of Gaussian kernels in the spatial block is greater than a first preset number; If the number is greater than the first preset number, the spatial block is orthogonally divided again until the number of Gaussian kernels in the spatial block is not greater than the first preset number. If it is not greater than, then determine whether the number of Gaussian kernels in the spatial block is less than the second preset number; wherein, the second preset number is less than the first preset number; If the number is less than the second preset number, the spatial block is merged with the adjacent spatial block until the number of Gaussian kernels in the spatial block is not less than the second preset number. The step of updating the parameters of the Gaussian kernels located in the field-of-view overlap region, the lens overlap region, and the dynamically changing region using the current frame panoramic image includes: Determine the target spatial blocks located in the field of view overlap region, the lens overlap region, and the dynamically changing region; The Gaussian kernel parameters in each of the target spatial blocks are updated using the current frame panoramic image.

9. The three-dimensional scene reconstruction method according to claim 8, characterized in that, The step of updating the Gaussian kernel parameters in each of the target spatial blocks using the current frame panoramic image includes: Based on the projection area of ​​the target spatial block, determine the original image in the projection area in the current frame panoramic image; Based on the pixel texture information of each pixel in the original image, set the weight value of each pixel; Based on the Gaussian kernel parameters of each Gaussian kernel, the Gaussian kernels in the target space block are projected onto the projection area to obtain a projection image; Calculate the brightness difference value of each pixel between the original image and the projected image, and determine the loss value using the brightness difference value and the weight value of each pixel; The parameters of each Gaussian kernel in the target space block are updated according to the loss value. When the parameter update is completed, the Gaussian kernel parameters of the abnormal Gaussian kernels that do not meet the preset parameter constraints are adjusted using the preset parameter constraints. Determine whether the iteration exit condition is met; wherein, the iteration exit condition is that the number of iterations reaches a first preset threshold, or the decrease in the loss value between two adjacent iterations is less than a second preset threshold; If the conditions are not met, proceed to the step of projecting the Gaussian kernels in the target space block onto the projection area according to the Gaussian kernel parameters of each Gaussian kernel to obtain the projected image; If the conditions are met, the parameter update for the Gaussian kernel is completed.

10. The three-dimensional scene reconstruction method according to claim 8, characterized in that, The step of updating the Gaussian kernel parameters in each of the target spatial blocks using the current frame panoramic image includes: Multiple target spatial blocks are distributed to multiple computing units, and each computing unit uses the current frame panoramic image to perform parallel iterative updates on the Gaussian kernel in each target spatial block.

11. The three-dimensional scene reconstruction method according to claim 1, characterized in that, Also includes: Gaussian kernels with transparency below a third preset threshold and three-dimensional scale less than a fourth preset threshold are filtered out, and it is determined whether the number of Gaussian kernels in the Gaussian kernel set is less than a fifth preset threshold. If it is smaller than the specified value, a Gaussian kernel is added to the overlapping field of view, the overlapping lens region, and the dynamically changing region.

12. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the three-dimensional scene reconstruction method as described in any one of claims 1 to 11.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when loaded and executed by a processor, implement the three-dimensional scene reconstruction method as described in any one of claims 1 to 11.

Images