A three-dimensional image reconstruction method using depth information

Abstract

The application relates to the technical field of image processing, in particular to a three-dimensional image reconstruction method using depth information, which comprises the following steps: collecting depth information of a target scene, generating a depth map based on the depth information of the target scene, and obtaining the depth value of each pixel point in the depth map corresponding to the distance of the pixel point from a camera in a real scene; in the target scene depth information collection stage, a target scene space model is obtained according to the boundary coordinates of the target scene; the application brings effective three-dimensional scene model modeling services for the depth map through the scene fast modeling technology of 3D Gaussian Splatting, in the method execution stage, the density of depth information sampling points is designed according to the complexity of the target scene space, the three-dimensional scene model modeling precision is effectively improved, and the depth map is subjected to prior denoising processing and smoothing processing after the three-dimensional scene model is constructed, so that the precision of the three-dimensional scene model constructed by the method is more comprehensively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and particularly relates to a three-dimensional image reconstruction method using depth information. BACKGROUND

[0002] Three-dimensional scene reconstruction is a process of converting two-dimensional images or point cloud data into a three-dimensional model. Through cameras, laser radars and other data collection, feature extraction, matching and depth calculation are performed to construct a point cloud. Then, a mesh model is generated by using techniques such as triangular meshing and surface reconstruction, and finally, texture mapping is performed to give the model a realistic appearance, which is widely used in games, buildings and other fields.

[0003] A three-dimensional image information acquisition device is disclosed in Chinese patent application No. 201510979644.4, which includes a mirror, a first area in the middle of the mirror, a first camera in the middle of the first area, a first light source on the mirror, the number of the first light source is more than four, which is distributed on the outer edge of the first area, the first camera is used to collect the light signal emitted by the first light source; the first light source and the first camera are connected with the control unit, the control unit is used to control the switching of each first light source at a predetermined time and control the working of the first camera at the corresponding time; the middle of the first area is also provided with a second camera, and more than one second light source is arranged on the outer edge of the first area, the second camera is used to collect the light signal emitted by the second light source; the first light source is an infrared LED lamp, and the first camera is an infrared camera; the second light source is a white light LED lamp, and the second camera is a visible light camera, the image reconstruction method includes: calibrating the intrinsic parameters of the first camera, and estimating the direction of the main optical axis of the first light source, so as to obtain the first light source coordinates; according to the intrinsic parameters of the first light source, a light emitting coordinate model is established.

[0004] The application aims to solve the problem that the existing photometric 3D reconstruction methods all adopt the assumption of point light source or parallel light source, that is, the light emitting characteristics of the light source itself are not considered, when the object is far away from the light source or the light source itself meets the parallel light or point light source characteristics, the existing method can obtain good 3D reconstruction effect, but the light source itself requires higher, and in order to obtain parallel light, the light source itself needs larger volume, such as using LED array device; it cannot realize the 3D image reconstruction at close distance, the overall volume of the device is large, the production cost is high, and the use is not convenient.

[0005] However, in the above-mentioned three-dimensional image reconstruction scene, there is no fine three-dimensional image reconstruction method through 3D Gaussian Splatting fast modeling technology.

[0006] To this end, a three-dimensional image reconstruction method using depth information is proposed. SUMMARY

[0007] In view of the above-mentioned defects of the prior art, the present application provides a three-dimensional image reconstruction method using depth information, which solves the technical problems proposed in the background art.

[0008] To achieve the above object, the present application is implemented by the following technical solutions:

[0009] A three-dimensional image reconstruction method using depth information, comprising:

[0010] Collecting target scene depth information, generating a depth map based on the target scene depth information, and obtaining the value of each pixel point in the depth map corresponding to its depth value from the camera in the real scene; in the target scene depth information collection stage, obtaining a target scene space model according to the target scene boundary coordinates, and deciding the target scene depth information collection point based on the complexity of the target scene space model, the collection angle of the target scene depth information in the collection stage is defined by the user terminal, and the collection is applied once at each collection point;

[0011] The complexity of the target scene space model is represented as:

[0012]

[0013] In the formula: C is the complexity of the target scene space model; N faces is the number of faces actually contained in the target scene space model; N ref is the number of reference faces; N edges is the number of edges actually possessed by the target scene space model; N ref-edges is the number of reference edges; N patches is the total number of discretized patches; k i is the average curvature of the i-th patch after discretization of the surface of the target scene space model; S ref-curvature is the total sum of curvature change values of the reference model surface; N high is the number of features presented by the target scene space model under high resolution; N low is the number of features that can be distinguished by the target scene space model under low resolution; N ref-detail is the difference in the number of features of the reference model under the same high and low resolution comparison, ω1, ω2, ω3, ω4 are weights, ω1, ω2, ω3, ω4 are all positive numbers and the sum is 1;

[0014] The depth map is acquired, the depth map is denoised, the depth map is converted from a camera coordinate system to a world coordinate system according to internal and external parameters of the camera, the depth map after the coordinate system conversion is completed is converted into three-dimensional point cloud data, an initial three-dimensional Gaussian distribution is associated with each point in the three-dimensional point cloud data, a mean vector, a covariance matrix and an amplitude parameter of each Gaussian distribution are determined, a range and an influence degree of each Gaussian distribution in space are identified to determine a weight and a spatial range of each three-dimensional Gaussian distribution, each three-dimensional Gaussian distribution is subjected to a splatting operation according to the weight and the spatial range, a visual three-dimensional scene model is generated, the three-dimensional scene model is further rendered, the three-dimensional scene model after the rendering is subjected to smoothing processing, and finally a reconstructed three-dimensional scene model is output.

[0015] Further, the target scene depth information includes distance information of an object and a camera, spatial position coordinates of the object, geometric shapes of an object contour, and terrain and topography sizes of a scene, the target scene depth information is collected by any one of a binocular vision depth camera, a structured light depth camera, a ToF camera, and a light field camera, a depth map generated based on the target scene depth information is output by the camera used for collecting the depth information, and a value of each pixel point in the depth map corresponds to a depth value of the pixel point from the camera in a real scene and is synchronously acquired by the camera used for collecting the depth information.

[0016] Further, the reference model is a cuboid model or a cube model defined by the user end, the number of features of the target scene space model is the number of concave-convex angles on a surface of the target scene space model, a ratio of a complexity C of the target scene space model defined by the user end to the number of target scene depth information collection points is defined by the user end, a product of the complexity C of the target scene space model and the ratio is the number of target scene depth information collection points, and a corresponding number of target scene depth information collection points are uniformly distributed on the target scene space model, so that adjacent intervals of the target scene depth information collection points are equal.

[0017] Further, the denoising logic of the depth map is represented as:

[0018]

[0019] In the formula, D filtered (i,j) is a depth value of a pixel point with coordinates (i,j) in the depth map after Gaussian filtering; k and l are sizes of a filtering window; G(m,n) is a Gaussian function value at a corresponding coordinate (m,n); and D(i+m,j+n) is an original depth value of a neighborhood pixel of the coordinate (i,j) defined by the filtering window.

[0020] wherein, σ is a standard deviation of the Gaussian function.

[0021] Furthermore, the camera's intrinsic and extrinsic parameters include focal length, optical center position, distortion coefficient, rotation matrix, and translation vector.

[0022] Furthermore, the operation of converting a depth map into 3D point cloud data is as follows:

[0023] Each pixel in the depth map is used as the processing target;

[0024]

[0025] In the formula: (X w Y w Z w (X) represents the three-dimensional coordinates of a pixel in the world coordinate system; c Y c Z c Let R be the 3D coordinates of the pixel in the camera coordinate system; and T be the rotation matrix and translation matrix, respectively.

[0026] In this process, after the three-dimensional coordinates of each pixel in the world coordinate system are obtained, the PointCloud class in the point cloud library is used to store each three-dimensional coordinate to obtain three-dimensional point cloud data.

[0027] Furthermore, the initial three-dimensional Gaussian distribution of the point associations in the three-dimensional point cloud data is as follows:

[0028] Points in 3D point cloud data include P w =(x w ,y w ,z w ), current point P w The coordinates are directly used as the mean vector μ of the associated three-dimensional Gaussian distribution. w =(x w ,y w ,z w );

[0029] The covariance matrix is ​​expressed as:

[0030] Where, σ x σ y σ z This represents the standard deviation along the corresponding coordinate axis.

[0031] The amplitude parameter is represented as: a w =ν×r w ν is the scaling factor set by the user based on experience; r w Let P be the point w The ratio of the point density in the neighborhood to the average point density of the entire point cloud.

[0032] Furthermore, when identifying the extent of the Gaussian distribution in space, the equiprobability contour of the three-dimensional Gaussian distribution is calculated based on the covariance matrix, and the calculation result is used to represent the range of the three-dimensional Gaussian distribution in space.

[0033] When identifying the influence of a 3D Gaussian distribution in space, the weight of the 3D Gaussian distribution during rendering or blending is set according to the amplitude parameter, and the setting result is used to represent the influence of the 3D Gaussian distribution in space.

[0034] in,

[0035] In the formula: ω w The weights are 3D Gaussian distributions used in rendering or blending; A w Let be the amplitude parameter of the w-th three-dimensional Gaussian distribution; G is the total number of three-dimensional Gaussian distributions.

[0036] Furthermore, when rendering the 3D scene model, the colors corresponding to each position in the target scene are applied for rendering. When the rendered 3D scene model is smoothed, smoothing is performed based on any one of the following algorithms: Laplacian smoothing algorithm, Taubin smoothing algorithm, and average curvature smoothing algorithm.

[0037] Furthermore, in the three-dimensional scene model output stage, the average distance error and root mean square error between the three-dimensional scene model and the real scene are calculated, an allowable error range is set, and the results are compared with the allowable error range. If the results meet the allowable error range, the output three-dimensional scene model is determined to be valid. If the results do not meet the allowable error range, the output three-dimensional scene model is determined to be invalid, and the reconstruction operation of the three-dimensional scene model corresponding to the target scene is re-executed.

[0038] In the process of reconstructing the 3D scene model corresponding to the target scene, the number of target scene depth information collection points used is greater than the number of target scene depth information collection points used in the previous 3D scene model reconstruction operation.

[0039] Compared with known public technologies, the technical solution provided by this invention has the following beneficial effects:

[0040] This invention provides a method for 3D image reconstruction using depth information. During the execution of this method, the method provides an effective 3D scene modeling service for the depth map through the rapid scene modeling technology of 3D Gaussian Splatting. In the execution stage, the density of depth information sampling points is designed considering the complexity of the target scene space, which effectively improves the modeling accuracy of the 3D scene model. Furthermore, the method performs prior denoising processing on the depth map and smoothing processing after the 3D scene model is constructed, which comprehensively improves the accuracy of the 3D scene model constructed by this method. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0042] Figure 1 This is a flowchart illustrating a three-dimensional image reconstruction method that utilizes depth information. Detailed Implementation

[0043] 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0044] The present invention will be further described below with reference to embodiments.

[0045] Example 1:

[0046] This embodiment presents a method for 3D image reconstruction using depth information, such as... Figure 1 As shown, it includes:

[0047] Collect depth information of the target scene, generate a depth map based on the depth information of the target scene, and obtain the value of each pixel in the depth map corresponding to its depth value from the camera in the real scene;

[0048] In the target scene depth information acquisition stage, the target scene spatial model is obtained based on the target scene boundary coordinates. The target scene depth information acquisition points are determined based on the complexity of the target scene spatial model. The acquisition perspective of the target scene depth information in the acquisition stage is customized by the user and is applied once at each acquisition point.

[0049] The complexity of the target scene space model is expressed as:

[0050]

[0051] In the formula: C represents the complexity of the target scene spatial model; N faces N represents the number of faces actually contained in the target scene spatial model. ref N is the reference face number. edges N represents the actual number of edges in the target scene space model; ref-edges N is the reference edge number; patches k represents the total number of discretized patches. i S represents the average curvature of the i-th facet after discretization of the surface of the target scene spatial model; ref-curvature N represents the sum of curvature changes on the surface of the reference model. high N represents the number of features presented by the target scene spatial model at high resolution. low N represents the number of features that the target scene spatial model can distinguish at low resolution; ref-detail The difference in the number of features of the reference model under the same high and low resolution comparison is represented by ω1, ω2, ω3, and ω4, which are weights. ω1, ω2, ω3, and ω4 are all positive numbers and their sum is 1.

[0052] The reference model is a user-defined cuboid or cube model. The number of features in the target scene space model is the number of concave and convex corners on the surface of the target scene space model. The ratio of the complexity C of the user-defined target scene space model to the number of target scene depth information collection points is defined by the user. The product of the complexity C of the target scene space model and the ratio is the number of target scene depth information collection points. The corresponding number of target scene depth information collection points are evenly distributed in the target scene space model, so that the adjacent spacing of each target scene depth information collection point is equal.

[0053] The complexity of the target scene spatial model is calculated using the above logical formula, and this is used as a reference to design the density of the target scene depth information collection points, ensuring that the construction of the target scene 3D model has complete data support.

[0054] Acquire a depth map, denoise the depth map, and transform the depth map from the camera coordinate system to the world coordinate system based on the camera's intrinsic and extrinsic parameters.

[0055] The denoising logic for depth maps is expressed as follows:

[0056]

[0057] In the formula: D filtered(i,j) is the depth value of the pixel at coordinates (i,j) in the depth map after Gaussian filtering; k and l are the sizes of the filtering window; G(m,n) is the Gaussian function value at the corresponding coordinates (m,n); D(i+m,j+n) is the original depth value of the neighboring pixels at coordinates (i,j) defined by the filtering window.

[0058] in, σ is the standard deviation of the Gaussian function;

[0059] The depth map is denoised using the above logical formula, thereby improving the accuracy of the final 3D scene model.

[0060] Obtain the depth map after coordinate system transformation and convert it into 3D point cloud data;

[0061] The operation to convert a depth map into 3D point cloud data is as follows:

[0062] Each pixel in the depth map is used as the processing target;

[0063]

[0064] In the formula: (X w Y w Z w (X) represents the three-dimensional coordinates of a pixel in the world coordinate system; c Y c Z c Let R be the 3D coordinates of the pixel in the camera coordinate system; and T be the rotation matrix and translation matrix, respectively.

[0065] The three-dimensional coordinates of each pixel in the world coordinate system are obtained and then stored using the PointCloud class in the point cloud library to obtain three-dimensional point cloud data.

[0066] For each point in the 3D point cloud data, associate it with an initial 3D Gaussian distribution and determine the mean vector, covariance matrix and magnitude parameter of each Gaussian distribution;

[0067] The initial 3D Gaussian distribution for point association in 3D point cloud data is:

[0068] Points in 3D point cloud data include P w =(x w ,y w ,z w ), current point P w The coordinates are directly used as the mean vector μ of the associated three-dimensional Gaussian distribution. w =(x w ,y w ,z w );

[0069] The covariance matrix is ​​expressed as: Where, σ x σ y σ z This represents the standard deviation along the corresponding coordinate axis.

[0070] The amplitude parameter is represented as: a w =ν×r w ν is the scaling factor set by the user based on experience; r w Let P be the point w The ratio of the point density in the neighborhood to the average point density of the entire point cloud;

[0071] Identify the spatial extent and degree of influence of each three-dimensional Gaussian distribution;

[0072] When identifying the spatial extent of a three-dimensional Gaussian distribution, the equiprobability contour of the three-dimensional Gaussian distribution is calculated based on the covariance matrix, and the calculation results are used to represent the range of the three-dimensional Gaussian distribution in space.

[0073] When identifying the influence of a 3D Gaussian distribution in space, the weight of the 3D Gaussian distribution during rendering or blending is set according to the amplitude parameter, and the setting result is used to represent the influence of the 3D Gaussian distribution in space.

[0074] in,

[0075] In the formula: ω w The weights are 3D Gaussian distributions used in rendering or blending; A w Let be the amplitude parameter of the w-th three-dimensional Gaussian distribution; G is the total number of three-dimensional Gaussian distributions;

[0076] Each 3D Gaussian distribution is splatting according to its weight and spatial range to generate a visualized 3D scene model. The 3D scene model is then rendered, and the rendered 3D scene model is simultaneously smoothed to finally output the reconstructed 3D scene model.

[0077] In this embodiment, a novel three-dimensional image reconstruction method is provided through the methods described in the above embodiments. Compared with the prior art, this method has controllable accuracy and a relatively stable and reliable reconstruction process.

[0078] Example 2:

[0079] At the implementation level, based on Example 1, this example refers to... Figure 1 The following is a further detailed description of a three-dimensional image reconstruction method using depth information in Example 1:

[0080] The target scene depth information includes: the distance between the object and the camera, the spatial coordinates of the object, the geometric shape of the object's outline, and the size of the scene's terrain. The target scene depth information is acquired by any one of the following: a binocular vision depth camera, a structured light depth camera, a ToF camera, or a light field camera. The depth map generated based on the target scene depth information is output by the camera used for depth information acquisition. The value of each pixel in the depth map corresponds to its depth value from the camera in the real scene, which is synchronously acquired by the camera used for depth information acquisition.

[0081] The above settings provide further execution data support for the execution of the method in Embodiment 1, ensuring the stable execution of the method in Embodiment 1.

[0082] like Figure 1 As shown, the camera's intrinsic and extrinsic parameters include focal length, optical center position, distortion coefficients, rotation matrix, and translation vector.

[0083] The above settings further limit the content of the camera's internal and external parameters.

[0084] like Figure 1 As shown, when rendering the 3D scene model, the colors corresponding to each position in the target scene are applied for rendering. When the rendered 3D scene model is smoothed, it is smoothed based on any one of the following algorithms: Laplacian smoothing algorithm, Taubin smoothing algorithm, and average curvature smoothing algorithm.

[0085] In the 3D scene model output stage, the average distance error and root mean square error between the 3D scene model and the real scene are calculated, and an error allowable range is set. Based on the error allowable range and the calculated results, if the calculated results meet the error allowable range, the output 3D scene model is determined to be valid. If the calculated results do not meet the error allowable range, the output 3D scene model is determined to be invalid, and the reconstruction operation of the 3D scene model corresponding to the target scene is re-executed.

[0086] In the process of reconstructing the 3D scene model corresponding to the target scene, the number of target scene depth information collection points used is greater than the number of target scene depth information collection points used in the previous 3D scene model reconstruction operation.

[0087] The above formula provides further optimization of the 3D scene model output by the method in Example 1, effectively improving the quality of the 3D scene model constructed by the method in Example 1.

[0088] In summary, the method in the above embodiments provides an effective 3D scene modeling service for the depth map through the rapid scene modeling technology of 3D Gaussian Splatting. During the method execution phase, the density of depth information sampling points is designed considering the complexity of the target scene space, which effectively improves the modeling accuracy of the 3D scene model. Furthermore, the method performs prior denoising processing on the depth map and smoothing processing after the 3D scene model is constructed, which comprehensively improves the accuracy of the 3D scene model constructed by this method.

[0089] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for reconstructing three-dimensional images using depth information, characterized in that, include: Collect depth information of the target scene, generate a depth map based on the depth information of the target scene, and obtain the value of each pixel in the depth map corresponding to its depth value from the camera in the real scene; Acquire a depth map, denoise the depth map, and transform the depth map from the camera coordinate system to the world coordinate system based on the camera's intrinsic and extrinsic parameters. Obtain the depth map after coordinate system transformation and convert it into 3D point cloud data; For each point in the 3D point cloud data, associate it with an initial 3D Gaussian distribution and determine the mean vector, covariance matrix, and magnitude parameter of each 3D Gaussian distribution; Identify the spatial extent and influence of each 3D Gaussian distribution to determine the weight and spatial extent of each 3D Gaussian distribution; Each 3D Gaussian distribution is splatting according to its weight and spatial range to generate a visualized 3D scene model. The 3D scene model is then rendered, and the rendered 3D scene model is simultaneously smoothed to finally output the reconstructed 3D scene model.

2. The method for three-dimensional image reconstruction using depth information according to claim 1, characterized in that, The target scene depth information includes: distance information between the object and the camera, spatial coordinates of the object, geometric shape of the object's outline, and the size of the scene's terrain. The target scene depth information is acquired by any one of binocular vision depth cameras, structured light depth cameras, ToF cameras, and light field cameras. The depth map generated based on the target scene depth information is output by the camera used for depth information acquisition. The value of each pixel in the depth map corresponds to its depth value from the camera in the real scene, which is synchronously acquired by the camera used for depth information acquisition.

3. The method for three-dimensional image reconstruction using depth information according to claim 1, characterized in that, In the target scene depth information acquisition stage, the target scene spatial model is obtained based on the target scene boundary coordinates. The target scene depth information acquisition point is determined based on the complexity of the target scene spatial model. The acquisition perspective of the target scene depth information in the acquisition stage is customized by the user and is applied once at each acquisition point. The complexity of the target scene spatial model is expressed as follows: In the formula: C represents the complexity of the target scene spatial model; N faces N represents the number of faces actually contained in the target scene spatial model. ref N is the reference face number. edges N represents the actual number of edges in the target scene space model; ref-edges N is the reference edge number; patches k represents the total number of discretized patches. i Let be the average curvature of the i-th facet after discretization of the surface of the target scene spatial model; S ref-curvature N represents the sum of curvature changes on the surface of the reference model. high N represents the number of features presented by the target scene spatial model at high resolution. low N represents the number of features that the target scene spatial model can distinguish at low resolution; ref-detail The difference in the number of features of the reference model under the same high and low resolution comparison is represented by ω1, ω2, ω3, and ω4, which are weights. ω1, ω2, ω3, and ω4 are all positive numbers and their sum is 1. The reference model is a user-defined cuboid or cube model. The number of features in the target scene space model is the number of concave and convex corners on the surface of the target scene space model. The ratio of the complexity C of the user-defined target scene space model to the number of target scene depth information collection points is defined by the user. The product of the complexity C of the target scene space model and the ratio is the number of target scene depth information collection points. The corresponding number of target scene depth information collection points are evenly distributed in the target scene space model, so that the adjacent spacing of each target scene depth information collection point is equal.

4. The method for three-dimensional image reconstruction using depth information according to claim 1, characterized in that, The denoising logic for the depth map is expressed as follows: In the formula: D filtered (i,j) is the depth value of the pixel at coordinates (i,j) in the depth map after Gaussian filtering; k and l are the sizes of the filtering window; G(m,n) is the Gaussian function value at the corresponding coordinates (m,n); D(i+m,j+n) is the original depth value of the neighboring pixels at coordinates (i,j) defined by the filtering window. in, σ is the standard deviation of the Gaussian function.

5. The method for three-dimensional image reconstruction using depth information according to claim 1, characterized in that, The camera's intrinsic and extrinsic parameters include focal length, optical center position, distortion coefficient, rotation matrix, and translation vector.

6. The method for three-dimensional image reconstruction using depth information according to claim 1, characterized in that, The operation to convert a depth map into 3D point cloud data is as follows: Each pixel in the depth map is used as the processing target; In the formula: (X w Y w Z w (X) represents the three-dimensional coordinates of a pixel in the world coordinate system; c Y c Z c Let R be the 3D coordinates of the pixel in the camera coordinate system; and T be the rotation matrix and translation matrix, respectively. In this process, after the three-dimensional coordinates of each pixel in the world coordinate system are obtained, the PointCloud class in the point cloud library is used to store each three-dimensional coordinate to obtain three-dimensional point cloud data.

7. The method for three-dimensional image reconstruction using depth information according to claim 1, characterized in that, The initial three-dimensional Gaussian distribution of the points associated in the three-dimensional point cloud data is as follows: Points in 3D point cloud data include P w =(x w ,y w ,z w ), current point P w The coordinates are directly used as the mean vector μ of the associated three-dimensional Gaussian distribution. w =(x w ,y w ,z w ); The covariance matrix is ​​expressed as: Where, σ x σ y σ z This represents the standard deviation along the corresponding coordinate axis. The amplitude parameter is represented as: a w =ν×r w ν is the scaling factor set by the user based on experience; r w Let P be the point w The ratio of the point density in the neighborhood to the average point density of the entire point cloud.

8. A three-dimensional image reconstruction method utilizing depth information according to claim 1, characterized in that, When identifying the spatial extent of a three-dimensional Gaussian distribution, the equiprobability contour of the three-dimensional Gaussian distribution is calculated based on the covariance matrix, and the calculation results are used to represent the range of the three-dimensional Gaussian distribution in space. When identifying the influence of a 3D Gaussian distribution in space, the weight of the 3D Gaussian distribution during rendering or blending is set according to the amplitude parameter, and the setting result is used to represent the influence of the 3D Gaussian distribution in space. in, In the formula: ω w The weights are 3D Gaussian distributions used in rendering or blending; A w Let be the amplitude parameter of the w-th three-dimensional Gaussian distribution; G is the total number of three-dimensional Gaussian distributions.

9. A three-dimensional image reconstruction method utilizing depth information according to claim 1, characterized in that, When rendering the 3D scene model, the colors corresponding to each position in the target scene are applied for rendering. When the rendered 3D scene model is smoothed, it is smoothed based on any one of the following algorithms: Laplacian smoothing algorithm, Taubin smoothing algorithm, and average curvature smoothing algorithm.

10. A three-dimensional image reconstruction method utilizing depth information according to claim 1, characterized in that, In the 3D scene model output stage, the average distance error and root mean square error between the 3D scene model and the real scene are calculated, and an error allowable range is set. Based on the error allowable range and the calculated result, if the calculated result meets the error allowable range, the output 3D scene model is determined to be valid. If the calculated result does not meet the error allowable range, the output 3D scene model is determined to be invalid, and the reconstruction operation of the 3D scene model corresponding to the target scene is re-executed. In the process of reconstructing the 3D scene model corresponding to the target scene, the number of target scene depth information collection points used is greater than the number of target scene depth information collection points used in the previous 3D scene model reconstruction operation.

Images