A three-dimensional scene reconstruction method and system based on SLAM and 3D Gaussian fusion

Abstract

The present invention provides a three-dimensional scene reconstruction method and system based on SLAM and 3D Gaussian fusion, including: step S1, shooting an indoor image for a target indoor scene and obtaining a dense point cloud A, and scanning to obtain a dense point cloud B; step S2, aligning the dense point cloud B to the dense point cloud A to obtain a fused point cloud; step S3, processing the fused point cloud to generate multiple Gaussian points, and recording the point cloud data points corresponding to each Gaussian point; step S4, based on each point cloud data point, aligning the normal of each Gaussian point with the normal of the fused point cloud; step S5, limiting the tangent vector distance between the center of each Gaussian point and the center of the fused point cloud; step S6, limiting the normal vector distance between the center of each Gaussian point and the center of the fused point cloud; step S7, obtaining a reconstructed three-dimensional scene for measuring the image distance between any two points in the three-dimensional scene and converting it into a real distance. The beneficial effect is that the present invention can improve the visual effect of the three-dimensional reconstruction scene.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and robot navigation, and in particular to a three-dimensional scene reconstruction method and system based on SLAM and 3D Gaussian fusion. Background Art

[0002] SLAM (simultaneous localization and mapping) technology originated in the field of robotics. It is an algorithm that allows devices (such as robots and drones) to locate themselves in unknown environments while building environmental maps. It uses data collected by sensors to achieve real-time environmental perception and self-positioning, and the point cloud collected by SLAM devices carries accurate and measurable location information. The core challenge of SLAM is that it requires both knowing the location to build a map and a map to determine the location. To solve this problem, SLAM uses probabilistic methods, especially Kalman filters, particle filters, or the recently popular graph optimization methods, to estimate the robot's trajectory and map features. SLAM has a wide range of applications in autonomous navigation, virtual reality, augmented reality, indoor positioning systems, and other fields.

[0003] 3D Gaussian Splatting (3DGS) is a 3D reconstruction technology that is mainly used to build high-quality 3D models from multi-view images. The core idea is to use Gaussian functions to represent each sampling point in the scene. Each Gaussian function carries color, transparency and spatial position information, which can effectively simulate the surface properties and volume information of real-world objects. 3DGS "projects" pixels in multi-view images onto Gaussian points in 3D space, and then fuses these points to construct a dense 3D representation. Compared with traditional volume rendering technology, 3DGS can better handle surface details and translucent effects, and show superior visual effects and higher reconstruction accuracy in work.

[0004] Although SLAM and 3DGS have certain advantages in 3D scene reconstruction, they still face the following problems:

[0005] 1. The 3D models built solely based on SLAM technology lack detail richness and visual quality;

[0006] 2. If the 3DGS algorithm is used alone during training, unreasonable Gaussian points will appear, resulting in overfitting problems and poor visual effects of the reconstructed scene;

[0007] 3. The 3DGS algorithm only relies on image information for 3D reconstruction and cannot obtain the precise distance in the scene. Summary of the invention

[0008] The technical problem to be solved by the present invention is to improve the visual effect of the three-dimensional reconstructed scene, avoid the generation of unreasonable point clouds, and achieve the acquisition of accurate distances in the scene. In order to overcome the defects of the above-mentioned prior art (or related technology), the present invention provides a three-dimensional scene reconstruction method and system based on slam and 3D Gaussian fusion.

[0009] The present invention provides a three-dimensional scene reconstruction method based on SLAM and 3D Gaussian fusion, comprising the following steps:

[0010] Step S1, photographing a preset target indoor scene to obtain an indoor image and obtaining a dense point cloud A based on the indoor image, and scanning the target indoor scene with a SLAM100 handheld laser scanner to obtain a dense point cloud B;

[0011] Step S2, aligning the dense point cloud B to the dense point cloud A to obtain a fused point cloud, and recording the deformation matrix when the dense point cloud B is aligned to the dense point cloud A;

[0012] Step S3, processing the fused point cloud based on the 3DGS algorithm to generate a plurality of Gaussian points, and recording a point cloud data point in the fused point cloud corresponding to each Gaussian point;

[0013] Step S4, based on each of the point cloud data points, aligning the normal of each of the Gaussian points with the normal of the fused point cloud through a normal vector loss function;

[0014] Step S5, limiting the tangent vector distance between the center of each Gaussian point and the center of the fused point cloud by a tangent vector distance loss function;

[0015] Step S6, limiting the normal vector distance between the center of each Gaussian point and the center of the fused point cloud by a normal vector distance loss function;

[0016] Step S7, training the fused point cloud based on the normal vector loss function, the tangent vector distance loss function and the normal vector distance loss function to obtain a reconstructed three-dimensional scene for measuring the image distance between any two points in the three-dimensional scene, and converting the image distance into a real distance in a real environment according to the deformation matrix.

[0017] Compared with the prior art, the three-dimensional scene reconstruction method based on SLAM and 3D Gaussian fusion of the present invention has the following advantages:

[0018] In the present invention, dense point cloud A and dense point cloud B are acquired through step S1, fused point cloud is generated through step S2, Gaussian points are generated through step S3, normal alignment of Gaussian points and fused point cloud is performed through step S4, tangent vector distance restriction of Gaussian points and fused point cloud is performed through step S5, normal vector distance restriction of Gaussian points and fused point cloud is performed through step S6, and three-dimensional scene is reconstructed through step S7. The dense point cloud B acquired from the SLAM100 handheld laser scanner is more accurate and can improve the visual effect of the three-dimensional reconstructed scene. The precise coordinate position of the dense point cloud B is used to guide the distribution and optimization of 3D Gaussian points, which can avoid unreasonable point cloud generation and ensure the realism and coherence of scene reconstruction. The reconstructed three-dimensional scene obtained by training based on the normal vector loss function, the tangent vector distance loss function and the normal vector distance loss function can be used to measure the image distance between any two points in the three-dimensional scene, and the accurate real distance in the real scene is obtained by transformation through the deformation matrix.

[0019] In a possible implementation, in the step S1, the dense point cloud A is obtained by importing the indoor image into metashape software for processing.

[0020] In a possible implementation, in step S2, a plurality of first position points are arbitrarily selected in the dense point cloud B, and a plurality of second position points are arbitrarily selected in the dense point cloud A, and every two of the first position points are used as a first position point group, and every two of the second position points are used as a second position point group. The first position point group and the second position point group at the same position are fused by Cloud Compare software to obtain the fused point cloud.

[0021] In a possible implementation, in step S4, the normal of each Gaussian point is aligned with the normal of the fused point cloud by the following calculation formula:

[0022] ;

[0023] in,

[0024] represents the normal vector loss function;

[0025] represents the number of Gaussian points;

[0026] represents the kth point cloud data point;

[0027] represents the kth Gaussian point;

[0028] represents the normal vector of the kth point cloud data point;

[0029] represents the normal vector of the k-th Gaussian point;

[0030] Represents the preset weight parameter.

[0031] In a possible implementation, in step S5, the tangent vector distance between the center of each Gaussian point and the center of the fused point cloud is limited by the following calculation formula:

[0032] ;

[0033] in,

[0034] represents the tangent vector distance loss function;

[0035] represents the number of Gaussian points;

[0036] represents the kth point cloud data point;

[0037] represents the kth Gaussian point;

[0038] represents the normal vector of the kth point cloud data point;

[0039] Indicates proximity The normal vectors of the 20 point cloud data points are The point where the angle difference between the normal vectors is the largest;

[0040] express and The L2 norm of .

[0041] In a possible implementation, in step S6, the normal vector distance between the center of each Gaussian point and the center of the fused point cloud is limited by the following calculation formula:

[0042] ;

[0043] in,

[0044] represents the normal vector distance loss function;

[0045] represents the number of Gaussian points;

[0046] represents the kth point cloud data point;

[0047] represents the kth Gaussian point;

[0048] Represents the normal vector of the kth point cloud data point.

[0049] The present invention also provides a three-dimensional scene reconstruction system based on SLAM and 3D Gaussian fusion, which is applied to the above-mentioned three-dimensional scene reconstruction method based on SLAM and 3D Gaussian fusion, comprising:

[0050] A point cloud acquisition module is used to obtain a dense point cloud A based on an indoor image pre-photographed for a target indoor scene, and to obtain a dense point cloud B of the target indoor scene scanned by a SLAM100 handheld laser scanner;

[0051] a point cloud fusion module, connected to the point cloud acquisition module, for aligning the dense point cloud B to the dense point cloud A to obtain a fused point cloud, and recording a deformation matrix when the dense point cloud B is aligned to the dense point cloud A;

[0052] a Gaussian point generation module, connected to the point cloud fusion module, for processing the fused point cloud based on a 3DGS algorithm to generate a plurality of Gaussian points, and recording a point cloud data point in the fused point cloud corresponding to each Gaussian point;

[0053] a normal alignment module, connected to the Gaussian point generation module, for aligning the normal of each Gaussian point with the normal of the fused point cloud through a normal vector loss function based on each point cloud data point;

[0054] a tangent quantity limiting module, connected to the Gaussian point generating module, for limiting the tangent vector distance between the center of each Gaussian point and the center of the fused point cloud through a tangent vector distance loss function;

[0055] a normal vector limiting module, connected to the Gaussian point generating module, and configured to limit the normal vector distance between the center of each Gaussian point and the center of the fused point cloud through a normal vector distance loss function;

[0056] A scene reconstruction module is respectively connected to the point cloud fusion module, the Gaussian point generation module, the normal alignment module, the tangent quantity limitation module and the normal vector limitation module, and is used to train the fused point cloud based on the normal vector loss function, the tangent vector distance loss function and the normal vector distance loss function to obtain a reconstructed three-dimensional scene for measuring the image distance between any two points in the three-dimensional scene.

[0057] In a possible implementation, the point cloud acquisition module obtains the dense point cloud A by importing the indoor image into metashape software for processing.

[0058] In a possible implementation, the point cloud fusion module arbitrarily selects multiple first position points in the dense point cloud B, and arbitrarily selects multiple second position points in the dense point cloud A, takes every two of the first position points as a first position point group, and takes every two of the second position points as a second position point group, and fuses the first position point group and the second position point group at the same position through Cloud Compare software to obtain the fused point cloud, and converts the image distance into the real distance in the real environment according to the deformation matrix. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is a flow chart of the method steps of the present invention;

[0060] Figure 2 It is the system structure principle diagram of the present invention;

[0061] Figure 3 A schematic diagram of measurement in a reconstructed three-dimensional scene of the present invention;

[0062] Description of reference numerals:

[0063] 1. Point cloud acquisition module; 2. Point cloud fusion module; 3. Gaussian point generation module; 4. Normal alignment module; 5. Tangent quantity limitation module; 6. Normal vector limitation module; 7. Scene reconstruction module. DETAILED DESCRIPTION

[0064] First, those skilled in the art should understand that these implementations are only used to explain the technical principles of the embodiments of the present invention, and are not intended to limit the protection scope of the embodiments of the present invention. Those skilled in the art can make adjustments to them as needed to adapt to specific application scenarios.

[0065] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] See also Figure 1 The embodiment of the present invention discloses a 3D scene reconstruction method based on SLAM and 3D Gaussian fusion, wherein point cloud information is obtained by SLAM technology and indoor images respectively, and then the point cloud information is fused using software such as Cloud Compare or Mesh Lab, and the fused point cloud information is processed using the 3DGS algorithm to generate Gaussian points with accurate distances for measuring the reconstructed 3D scene. The technical solutions adopted are as follows:

[0067] Step S1, photographing a preset target indoor scene to obtain an indoor image and obtaining a dense point cloud A based on the indoor image, and scanning the target indoor scene with a SLAM100 handheld laser scanner to obtain a dense point cloud B;

[0068] Step S2, aligning the dense point cloud B to the dense point cloud A to obtain a fused point cloud, and recording the deformation matrix when the dense point cloud B is aligned to the dense point cloud A;

[0069] Step S3, processing the fused point cloud based on the 3DGS algorithm to generate multiple Gaussian points, and recording the point cloud data point in the fused point cloud corresponding to each Gaussian point;

[0070] Step S4, based on each point cloud data point, aligning the normal of each Gaussian point with the normal of the fused point cloud through a normal vector loss function;

[0071] Step S5, limiting the tangent vector distance between the center of each Gaussian point and the center of the fused point cloud by using a tangent vector distance loss function;

[0072] Step S6, limiting the normal vector distance between the center of each Gaussian point and the center of the fused point cloud by a normal vector distance loss function;

[0073] Step S7, training the fused point cloud based on the normal vector loss function, the tangent vector distance loss function and the normal vector distance loss function to obtain a reconstructed three-dimensional scene for measuring the image distance between any two points in the three-dimensional scene, and converting the image distance into the real distance in the real environment according to the deformation matrix.

[0074] Continue to see Figure 1 ,In step S1, after acquiring the indoor image, a dense point cloud A is obtained using software such as ,metashape.

[0075] Continue to see Figure 1 In step S2, the dense point cloud B generated by the SLAM100 handheld laser scanner is aligned to the dense point cloud A through software such as Cloud Compare, and the deformation matrix of the dense point cloud B to the dense point cloud A is recorded. The detailed process is shown in Figure 3 , in the dense point cloud B generated by the SLAM100 handheld laser scanner, arbitrarily select two points A0 and A1, and in the dense point cloud A generated by metashape, select the corresponding two points R0 and R1. The two point clouds can be fused through Cloud Compare to obtain a fused point cloud.

[0076] Continue to see Figure 1The SLAM100 handheld laser scanner has a paired software. After scanning, the device data can be processed by the software to obtain a dense point cloud. The present invention is not limited to the SLAM100 handheld laser scanner. Any SLAM device that can export point cloud information with precise position can be used.

[0077] Continue to see Figure 1 Because the dense point cloud A given by metashape is different from the dense point cloud B given by SLAM100, their spatial coordinates, sizes, rotations, etc. are all different. Therefore, the alignment in step S2 refers to overlapping the position of each point in the dense point cloud B given by SLAM100 with the dense point cloud A given by metashape. This is achieved using software such as cloud compare.

[0078] Continue to see Figure 1 ,In step S3, the Gaussian points generated by the fused point cloud are trained by the improved 3DGS algorithm, so that the Gaussian points match the fused point cloud, and the point cloud data points corresponding to each Gaussian point are recorded.

[0079] Continue to see Figure 1 In step S4, the normal line of the Gaussian point is aligned with the normal line of the fused point cloud through the normal vector loss function, as shown in the following formula:

[0080] ;

[0081] in, is the point cloud data point of the kth dense point cloud B, is the kth Gaussian point, is the normal vector of the kth point cloud data point, is the normal vector of the kth Gaussian point, is a weight parameter, which is calculated by taking the 20 closest The standard deviation of the normal vector of the point cloud data point is calculated. The more regular the surface of the object is, the The bigger the smaller, and vice versa.

[0082] Continue to see Figure 1 In step S5, the tangent vector distance loss function is used to limit the tangent vector distance between the center of the Gaussian point and the center of the fused point cloud, as shown in the following formula:

[0083] ;

[0084] in, is the point cloud data point of the kth dense point cloud B, Is adjacent The normal vectors of the 20 point cloud data points are The point with the largest difference in normal vector angles, yes and The L2 norm of the tangent vector distance loss function can be used to control the distance between the center of the Gaussian point and the center of the dense point cloud B to remain close, so that the spatiality is maintained and the geometric shape of the object in the reconstructed three-dimensional scene is maintained.

[0085] Continue to see Figure 1 In step S6, the normal vector distance loss function is used to limit the normal vector distance between the center of the Gaussian point and the center of the fused point cloud, as shown in the following formula:

[0086] ;

[0087] Through the above normal vector distance loss function, the distance between the center of the Gaussian point and the center of the dense point cloud B can be controlled to remain close, so that the spatiality is maintained, and at the same time, the geometric shape of the object in the reconstructed three-dimensional scene is maintained.

[0088] Continue to see Figure 1 In step S7, the loss function in steps S4, S5 and S6 is used for training to obtain a reconstructed three-dimensional scene. Since the Gaussian during training is generated by a dense point cloud B with position information, there will be no floating objects in the reconstructed three-dimensional scene, and the visual effect is good. By finding any two points A0 and A1 in the three-dimensional scene and measuring the distance between the two points, the distance between the two points in the real environment can be obtained by using the deformation matrix in the above steps. After calculation, the measured distance obtained in the reconstructed three-dimensional scene is 156.71cm, and the distance measured in the real environment is 158.23cm. The relative error of the measured distance is 0.96%, which meets the actual engineering needs.

[0089] See also Figure 2 The embodiment of the present invention further discloses a three-dimensional scene reconstruction system based on SLAM and 3D Gaussian fusion, comprising:

[0090] A point cloud acquisition module 1 is used to obtain a dense point cloud A based on an indoor image pre-photographed for a target indoor scene, and to obtain a dense point cloud B of the target indoor scene scanned by a SLAM100 handheld laser scanner;

[0091] A point cloud fusion module 2 is connected to the point cloud acquisition module 1 and is used to align the dense point cloud B to the dense point cloud A to obtain a fused point cloud, and record the deformation matrix when the dense point cloud B is aligned to the dense point cloud A;

[0092] A Gaussian point generation module 3, connected to the point cloud fusion module 2, is used to process the fused point cloud based on the 3DGS algorithm to generate multiple Gaussian points, and record the point cloud data point corresponding to each Gaussian point in the fused point cloud;

[0093] a normal alignment module 4, connected to the Gaussian point generation module 3, for aligning the normal of each Gaussian point with the normal of the fused point cloud through a normal vector loss function based on each point cloud data point;

[0094] A tangent quantity limiting module 5, connected to the Gaussian point generating module 3, is used to limit the tangent vector distance between the center of each Gaussian point and the center of the fused point cloud through a tangent vector distance loss function;

[0095] a normal vector limiting module 6, connected to the Gaussian point generating module 3, for limiting the normal vector distance between the center of each Gaussian point and the center of the fused point cloud through a normal vector distance loss function;

[0096] A scene reconstruction module 7 is respectively connected to the point cloud fusion module 2, the Gaussian point generation module 3, the normal alignment module 4, the tangent quantity limitation module 5 and the normal vector limitation module 6, and is used to train the fused point cloud based on the normal vector loss function, the tangent vector distance loss function and the normal vector distance loss function to obtain a reconstructed three-dimensional scene for measuring the image distance between any two points in the three-dimensional scene, and convert the image distance into a real distance in a real environment according to the deformation matrix.

[0097] Continue to see Figure 2 ,The point cloud acquisition module 1 obtains a dense point cloud A by importing the indoor image into the metashape software.

[0098] Continue to see Figure 2 The point cloud fusion module 2 arbitrarily selects multiple first position points in the dense point cloud B, and arbitrarily selects multiple second position points in the dense point cloud A, takes every two first position points as a first position point group, and takes every two second position points as a second position point group, and fuses the first position point group and the second position point group at the same position through the Cloud Compare software to obtain a fused point cloud.

[0099] In the description of the present invention, the description with reference to the terms "one embodiment", "some embodiments", "in the present embodiment", "specific example", or "some examples" etc. means that the specific features, mechanisms, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, mechanisms, materials or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0100] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A three-dimensional scene reconstruction method based on SLAM and 3D Gaussian fusion, characterized in that: The following steps are involved: Step S1, photographing a preset target indoor scene to obtain an indoor image and obtaining a dense point cloud A based on the indoor image, and scanning the target indoor scene with a SLAM100 handheld laser scanner to obtain a dense point cloud B; Step S2, aligning the dense point cloud B to the dense point cloud A to obtain a fused point cloud, and recording the deformation matrix when the dense point cloud B is aligned to the dense point cloud A; Step S3, processing the fused point cloud based on the 3DGS algorithm to generate a plurality of Gaussian points, and recording a point cloud data point in the fused point cloud corresponding to each Gaussian point; Step S4, based on each of the point cloud data points, aligning the normal of each of the Gaussian points with the normal of the fused point cloud through a normal vector loss function; Step S5, limiting the tangent vector distance between the center of each Gaussian point and the center of the fused point cloud by a tangent vector distance loss function; Step S6, limiting the normal vector distance between the center of each Gaussian point and the center of the fused point cloud by a normal vector distance loss function; Step S7, training the fused point cloud based on the normal vector loss function, the tangent vector distance loss function and the normal vector distance loss function to obtain a reconstructed three-dimensional scene for measuring the image distance between any two points in the three-dimensional scene, and converting the image distance into a real distance in a real environment according to the deformation matrix.

2. The three-dimensional scene reconstruction method based on SLAM and 3D Gaussian fusion according to claim 1, characterized in that: In the step S1, the dense point cloud A is obtained by importing the indoor image into metashape software for processing.

3. The three-dimensional scene reconstruction method based on SLAM and 3D Gaussian fusion according to claim 1, characterized in that: In the step S2, a plurality of first position points are arbitrarily selected in the dense point cloud B, and a plurality of second position points are arbitrarily selected in the dense point cloud A, and every two of the first position points are used as a first position point group, and every two of the second position points are used as a second position point group. The first position point group and the second position point group at the same position are fused by Cloud Compare software to obtain the fused point cloud.

4. The three-dimensional scene reconstruction method based on SLAM and 3D Gaussian fusion according to claim 1, characterized in that: In step S4, the normal of each Gaussian point is aligned with the normal of the fused point cloud by the following calculation formula: ； in, represents the normal vector loss function; represents the number of Gaussian points; represents the kth point cloud data point; represents the kth Gaussian point; represents the normal vector of the kth point cloud data point; represents the normal vector of the k-th Gaussian point; Represents the preset weight parameter.

5. The three-dimensional scene reconstruction method based on SLAM and 3D Gaussian fusion according to claim 1, characterized in that: In step S5, the tangent vector distance between the center of each Gaussian point and the center of the fused point cloud is limited by the following calculation formula: ； in, represents the tangent vector distance loss function; represents the number of Gaussian points; represents the kth point cloud data point; represents the kth Gaussian point; represents the normal vector of the kth point cloud data point; Indicates proximity The normal vectors of the 20 point cloud data points are The point where the angle difference between the normal vectors is the largest; express and The L2 norm of .

6. The three-dimensional scene reconstruction method based on SLAM and 3D Gaussian fusion according to claim 1, characterized in that: In step S6, the normal vector distance between the center of each Gaussian point and the center of the fused point cloud is limited by the following calculation formula: ； in, represents the normal vector distance loss function; represents the number of Gaussian points; represents the kth point cloud data point; represents the kth Gaussian point; Represents the normal vector of the kth point cloud data point.

7. A three-dimensional scene reconstruction system based on SLAM and 3D Gaussian fusion, characterized in that: The method for reconstructing a three-dimensional scene based on SLAM and 3D Gaussian fusion as claimed in any one of claims 1 to 6 comprises: A point cloud acquisition module (1) is used to obtain a dense point cloud A based on an indoor image pre-photographed for a target indoor scene, and to obtain a dense point cloud B of the target indoor scene scanned by a SLAM100 handheld laser scanner; A point cloud fusion module (2), connected to the point cloud acquisition module (1), used for aligning the dense point cloud B with the dense point cloud A to obtain a fused point cloud, and recording a deformation matrix when the dense point cloud B is aligned with the dense point cloud A; a Gaussian point generation module (3), connected to the point cloud fusion module (2), for processing the fused point cloud based on a 3DGS algorithm to generate a plurality of Gaussian points, and recording a point cloud data point in the fused point cloud corresponding to each Gaussian point; a normal alignment module (4), connected to the Gaussian point generation module (3), for aligning the normal of each Gaussian point with the normal of the fused point cloud based on each point cloud data point through a normal vector loss function; a tangent quantity limiting module (5), connected to the Gaussian point generating module (3), and used for limiting the tangent vector distance between the center of each Gaussian point and the center of the fused point cloud through a tangent vector distance loss function; a normal vector limiting module (6), connected to the Gaussian point generating module (3), and used for limiting the normal vector distance between the center of each Gaussian point and the center of the fused point cloud through a normal vector distance loss function; A scene reconstruction module (7) is respectively connected to the point cloud fusion module (2), the Gaussian point generation module (3), the normal alignment module (4), the tangent quantity limitation module (5) and the normal vector limitation module (6), and is used to train the fused point cloud based on the normal vector loss function, the tangent vector distance loss function and the normal vector distance loss function to obtain a reconstructed three-dimensional scene for measuring the image distance between any two points in the three-dimensional scene.

8. The three-dimensional scene reconstruction system based on SLAM and 3D Gaussian fusion according to claim 7, characterized in that: The point cloud acquisition module (1) obtains the dense point cloud A by importing the indoor image into metashape software for processing.

9. The three-dimensional scene reconstruction system based on SLAM and 3D Gaussian fusion according to claim 7, characterized in that: The point cloud fusion module (2) randomly selects a plurality of first position points in the dense point cloud B, randomly selects a plurality of second position points in the dense point cloud A, takes every two of the first position points as a first position point group, takes every two of the second position points as a second position point group, fuses the first position point group and the second position point group at the same position through Cloud Compare software to obtain the fused point cloud, and converts the image distance into a real distance in a real environment according to the deformation matrix.

Images