A gauss representation slam method based on dense matching prior and factor graph constraint

Abstract

The application discloses a kind of Gauss representation SLAM methods based on dense matching prior and factor graph constraint, including input current image and key frame image, pre-trained model, output the point graph corresponding to image pair, and return the matching situation of two frame image points and respective point cloud information, based on point level matching result, with photometric consistency error and geometric projection error as target, construct the joint optimization problem of current frame and key frame, jointly estimate the camera pose and point cloud of current frame, realize high-precision pose solution, after Gauss scene representation and rasterization rendering processing, generate point graph data, and pass to back end and carry out global optimization, back end receives pose and point cloud data, executes loop detection to identify repeated key frame, Gauss rendering is carried out through optimized key frame graph, complete global dense three-dimensional reconstruction, the present application effectively solves the trajectory drift and scene inconsistency problem caused by lack of pose prior and global geometric constraint in existing system.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and robot navigation technology, specifically to a Gaussian representation SLAM method based on dense matching priors and factor graph constraints. Background Technology

[0002] Simultaneous Localization and Mapping (SLAM) aims to estimate its own pose and reconstruct a 3D scene in real time using camera sensors in unknown environments. Traditional sparse or semi-dense visual SLAM methods rely on manually designed features or image gradients for matching and geometric optimization. While computationally efficient and with good localization accuracy, the generated maps are typically sparse, lack continuity and detail, and are susceptible to tracking loss due to changes in lighting or weakly textured environments. To address these issues, deep learning-based dense SLAM extracts features and predicts dense depth through neural networks, achieving robust localization in complex environments. However, its depth prediction suffers from scale drift and noise accumulation, leading to poor point cloud consistency and significant global errors.

[0003] To improve the continuity of reconstruction and rendering quality, researchers have introduced neural scene representation into the SLAM framework, achieving differentiable 3D modeling through implicit function learning. A representative method is Neural Radiation Field (NeRF). NeRF achieves continuously differentiable 3D scene representation through implicit volume functions, and can jointly optimize camera pose and radiation parameters to generate high-fidelity rendered images. Its advantages include strong geometric continuity and high rendering quality, but it suffers from slow training and inference speeds, high dependence on pose accuracy, and high computational and storage overhead, making it difficult to meet the requirements of real-time SLAM systems. 3D Gaussian Splashing (3D-GS) explicitly represents the scene structure using a discrete Gaussian distribution, combined with rasterization rendering to achieve real-time differentiable 3D reconstruction. This method has advantages such as fast rendering speed, high storage efficiency, and dynamic updates, but it still relies on accurate pose input, and front-end errors can easily lead to drift. In large-scale scenes, the number of Gaussians surges, and the lack of a global optimization mechanism makes it difficult to maintain long-term consistency. Summary of the Invention

[0004] The purpose of this invention is to provide a Gaussian representation SLAM method based on dense matching priors and factor graph constraints to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a Gaussian representation SLAM method based on dense matching priors and factor graph constraints, comprising the following operational steps:

[0006] Step S1: Input the current image and keyframe images, use the pre-trained Vision Transformer model to extract feature points and construct point clouds, output the point map corresponding to the image, and return the matching status of the points of the two frames and their respective point cloud information.

[0007] Preferably, the current frame image and a reference keyframe image are received as input, and a pre-trained multimodal visual transformation model is used for feature extraction and spatial mapping to generate a 3D point map corresponding to each frame. The model establishes a point-level correspondence between frames and outputs the matching index, matching confidence, and geometric and photometric information of the point clouds of the two frames, providing initial constraints for subsequent pose estimation.

[0008] Step S2: Based on the point-level matching results, high-precision pose estimation optimization of the current frame is obtained through ray constraint optimization. Using the optimized relative pose, the point map of the key frame is spatially transformed and fused to update, a locally consistent point cloud map is constructed, Gaussian scene representation and rasterization rendering are performed, visualization reconstruction and secondary pose correction are completed, and point map data with high-precision pose is generated and transmitted to the backend.

[0009] Preferably, pose optimization based on ray distance constraints aims to achieve fast pose tracking between the current frame and keyframes. A low-latency ray constraint fusion tracking and point cloud filtering method is proposed, which obtains the pose by minimizing 3D point errors. However, this method is susceptible to point cloud noise when depth estimation errors exist. Therefore, given the camera intrinsic parameters, an optimization model based on ray direction error is used to replace the depth-sensitive point error term. After converting the point map into unit direction vector form, ray direction difference constraints are applied to each pair of matching points. During the solution process, a Gauss-Newton iteration and iterative reweighted least squares framework are introduced to calculate the analytical Jacobian matrix of pose perturbation caused by ray and distance residuals. By constructing the residual matrix, Jacobian matrix, weight matrix, and relative transformation matrix, the linear equation is iteratively solved to update the pose.

[0010] Preferably, a high-precision pose estimate for the current frame is obtained through ray constraint optimization. Subsequently, the optimized relative pose is used to perform spatial transformation and fusion updates on the point map of the keyframe to achieve a locally consistent point cloud map construction. After Gaussian scene representation and rasterization rendering, further visualization reconstruction and secondary pose correction can be performed to generate point map data with high-precision pose and transmit it to the backend.

[0011] Step S3: The backend receives pose and point cloud data, performs loop closure detection, identifies repeated keyframes, introduces global constraints and performs second-order nonlinear global optimization when a loop closure is detected, and establishes local constraints to maintain map continuity and scale consistency when no loop closure is detected. Gaussian rendering is then performed using the optimized keyframe map to complete the global dense 3D reconstruction.

[0012] Preferably, the backend first manages the keyframe sequence output by the frontend. When the number of valid matches or the number of unique keyframe pixels between consecutive frames falls below a threshold, the system automatically adds a new keyframe and establishes a bidirectional edge between it and the previous keyframe to add it to the edge set. This constraint ensures the continuity of adjacent keyframes in the time series, but pose drift may still occur due to the existence of accumulated errors; therefore, an aggregation-based selection matching kernel is introduced.

[0013] An incremental loop closure detection mechanism (ASMK) is employed. By jointly implementing incremental loop closure detection and global optimization of ray constraints, global consistency maintenance between keyframes is achieved, effectively eliminating front-end cumulative drift and ensuring long-term stable operation of the system in large-scale scenes. Utilizing analytical Jacobian matrices and parallel reduction to accelerate the construction of Hessian matrices using CUDA, the back-end optimization of this invention can achieve efficient global pose and geometric consistency correction without increasing system bottlenecks. Compared to traditional first-order methods, the second-order information introduced in this method significantly improves convergence speed and global optimization quality.

[0014] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention combines dense matching priors with factor graph backend optimization, and introduces a closed-loop mechanism of incremental loop retrieval and dual-input Transformer decoder geometric verification. Simultaneously, it employs a second-order global optimization strategy of Gaussian-Newton iteration and sparse Cholesky decomposition to jointly constrain and correct the consistency of keyframe pose and scene geometry. This method achieves dense, high-fidelity 3D reconstruction and high-precision pose estimation while ensuring real-time performance, effectively solving the trajectory drift and scene inconsistency problems caused by the lack of pose priors and global geometric constraints in existing Gaussian representation-based SLAM systems. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0016] Figure 1 This is a flowchart illustrating the Gaussian representation SLAM method based on dense matching prior and factor graph constraints provided in this embodiment of the invention.

[0017] Figure 2 This is a schematic diagram of the loop closure detection process based on the aggregation selection matching kernel ASMK provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Combination Figures 1-2 As shown, the present invention provides a technical solution: a Gaussian representation SLAM method based on dense matching prior and factor graph constraints, comprising the following steps:

[0020] Step S1: Input the current image and keyframe image, and output the point map corresponding to the image after the pre-trained model, and return the matching situation of the points of the two frames and their respective point cloud information.

[0021] In this embodiment, the system receives the current frame image and a reference keyframe image as input, and calls a pre-trained multimodal visual transformation model to perform feature extraction and spatial mapping. This model integrates a dual-input Transformer decoder structure to achieve cross-view feature fusion and bidirectional point cloud prediction, thereby outputting a 3D point map corresponding to each frame. The system calculates the pixel-level correspondence between two frames through a matching module, outputting the matching index, matching validity mask, and respective geometric, confidence, and feature descriptor information between the two frames. Through the above process, the system obtains dense matching and high-confidence geometric information between the two frames, providing direct geometric constraints for subsequent relative pose estimation.

[0022] Step S2: Based on the point-level matching results, the system constructs a joint optimization problem between the current frame and keyframes, targeting photometric consistency error and geometric projection error, and jointly estimates the camera pose and point cloud of the current frame. During the optimization process, the Sim(3) model based on ray distance constraints or the SE(3) model based on intrinsic parameter constraints is called to achieve high-precision pose solving. After Gaussian scene representation and rasterization rendering, point map data with high-precision pose is generated and passed to the backend module for global optimization.

[0023] In this embodiment, pose optimization based on ray distance constraints aims to achieve rapid pose tracking between the current frame and keyframes. This invention proposes a low-latency ray constraint fusion tracking and point cloud filtering method. This method obtains the pose by minimizing 3D point errors, but it is easily affected by point cloud noise when depth estimation errors exist. Therefore, when the camera intrinsic parameters are known, an optimization model based on ray direction error is used to replace the depth-sensitive point error term. After converting the point map into a unit direction vector form, ray direction difference constraints are applied to each pair of matching points.

[0024] For example, the system estimates the current frame based on keyframes. Compared to the previous keyframe The relative transformation matrix between Assuming the point map estimation of the previous keyframe is known Where the subscript k indicates that the point originates from a keyframe. The superscript k indicates that the point is in the keyframe. In the camera coordinate system, this can be expressed through the feature matching module. Obtain the 3D points corresponding to the current frame Used to solve pose transformation When the camera intrinsic parameters are unknown, pose estimation is performed using a scale-consistent ray distance constraint:

[0025]

[0026] in, This represents the point cloud matching error. This represents a scale-consistent similarity transformation used for pose estimation in uncalibrated scenes. For feature matching confidence weights, ,in, Indicates the current frame The Middle The matching confidence of each point Keyframe The Middle The matching confidence of each point To match the expected variance of the residuals, For robust nuclear norm, its corresponding loss function Huber or Cauchy forms can be used to suppress the impact of outlier matching points on the optimization results. Keyframe The first in its own coordinate system Three-dimensional points, weight function As shown in the formula:

[0027]

[0028] Given the camera intrinsic parameters, pose optimization based on ray direction error is shown in the formula:

[0029]

[0030] in, This represents the error in the direction of light rays. When camera intrinsics are enabled, the error in the direction of light rays is used. The constraints are pure rotation and translation. This indicates a direction normalization operation. The residual noise variance is a strong robust constraint on depth prediction error and can effectively suppress attitude drift caused by local geometric inconsistencies. Since the angle error is bounded, the above equation can still maintain stable convergence when there are outliers.

[0031] Furthermore, after each pose calculation, a weighted average filtering strategy is used to update the point map of the keyframes to fuse observation information from different perspectives, thereby enhancing geometric consistency. The updated model is shown in the figure.

[0032]

[0033] in: Keyframe The current fusion point map (expressed in its own coordinate system); For the current frame 3D point map, This represents the result of its transformation to the keyframe coordinate system; This is the fused confidence map corresponding to the keyframe point map, representing the accumulated observation confidence. This is the confidence map of the current frame in the keyframe coordinate system. The symbol "←" indicates an assignment operation (update), that is, updating the variables on the left with the calculation results on the right. This weighted average filter can continuously integrate geometric information from multiple frames without global optimization, effectively improving the consistency and robustness of the point map. In addition, this fusion strategy considers both geometric estimation and confidence weights, thereby effectively suppressing accumulated errors in scenarios with depth noise while maintaining real-time performance and robustness.

[0034] For example, after obtaining the optimized camera pose parameters, the system further introduces a 3D Gaussian scene representation to achieve differentiable photometric consistency optimization. This Gaussian representation is used to construct a continuous volumetric model of the scene, enabling the rendering result to align with the real image at the pixel level. The photometric loss function is defined as follows:

[0035]

[0036] in, Indicates loss of photometric uniformity. Representation based on Gaussian sets With camera pose The pixel-by-pixel differentiable rendering function is used to generate rendered images. Represents a real image.

[0037] For example, in the solution process, a Gauss-Newton iteration and iterative reweighted least squares framework is introduced to calculate the effect of ray and distance residuals on pose perturbation. The analytical Jacobian matrix is ​​obtained by constructing the residual matrix. Jacobian matrix and weight matrix relative transformation matrix The linear equation is solved iteratively and the pose is updated, as shown in the following formula:

[0038]

[0039] The symbol ⊕ indicates that the pose perturbation The exponential mapping is applied to the camera pose transformation matrix. superior.

[0040] In this embodiment, a high-precision pose estimation of the current frame is obtained through ray constraint optimization. Then, the optimized relative pose is used to perform spatial transformation and fusion update of the point map of the key frame to achieve the construction of a locally consistent point cloud map. After processing by Gaussian scene representation and rasterization rendering, further visualization reconstruction and secondary pose correction can be performed to generate point map data with high-precision pose and transmit it to the backend.

[0041] Step S3: The backend receives the pose and point cloud data from the frontend and performs loop closure detection to identify repeating keyframes. When a loop closure is detected, global constraints are introduced and second-order nonlinear global optimization is performed; when no loop closure is detected, local constraints are established to maintain map continuity and scale consistency. Finally, Gaussian rendering is performed on the optimized keyframe map to complete the globally dense 3D reconstruction.

[0042] In this embodiment, the backend module receives pose optimization results and point cloud data from the frontend and performs loop closure detection to identify keyframe pairs that are observed repeatedly, thereby achieving globally consistent pose and map optimization. The backend first manages the keyframe sequence output by the frontend. When the number of valid matches between consecutive frames or the number of unique keyframe pixels is below a threshold... At that time, the system automatically adds new keyframes. and in its connection to the previous keyframe Establish bidirectional edges between them to add to the edge set. This constraint ensures the continuity of adjacent keyframes in the time series, but pose drift may still occur due to the existence of accumulated errors.

[0043] For example, this invention introduces an incremental loop closure detection mechanism based on the Aggregate Selection Matching Kernel (ASMK). By combining incremental loop closure detection with global optimization of ray constraints, global consistency maintenance between keyframes is achieved, effectively eliminating front-end cumulative drift and ensuring long-term stable operation of the system in large-scale scenarios.

[0044] Specifically, using the current keyframe The encoded features are queried in the database to retrieve the top results with the highest similarity. Several candidate keyframes. Since the encoded features use clustered quantization (containing only tens of thousands of cluster centers), features are assigned to the nearest cluster center by calculating the L2 distance (Euclidean distance) between the feature vector and each cluster center, achieving fast feature quantization. When the retrieval score exceeds a set threshold... At that time, candidate keyframe pairs are input into the dual-input Transformer decoder module for dense feature matching; if the number of matched feature points exceeds a threshold... Then, bidirectional constraint edges are added to the frame pair in the keyframe graph to form a closed-loop constraint. Finally, the encoded features of the new keyframe are inserted into the inverse file index structure to dynamically update the retrieval database, achieving online updates and rapid response for loop closure detection.

[0045] After loop closure detection is completed, the system enters the backend global optimization phase. This involves estimating the current pose of a given keyframe. and its corresponding standard point map The objective of this invention is to achieve globally consistent optimization of the pose and geometry of all keyframes. In the case of unknown camera intrinsic parameters, this optimization is based on a global error model constrained by light, as shown in the formula:

[0046]

[0047] in, Represents the global geometric error term. Keyframe and Relative pose transformation between them express, This indicates that the weighted terms Confidence weights for feature matching With noise variance Together, to enhance the impact of high-confidence matching and reduce outlier interference, given the camera intrinsic parameters, this optimization is based on a global error model with pixel projection constraints, as shown in the formula:

[0048]

[0049] in, This represents the global reprojection error based on pixel projection. This is a pixel projection function based on a known camera model. Indicates the first The pixel coordinates of the frame correspond to the point. This represents the projection noise variance. The model improves geometric consistency accuracy through pixel-domain error constraints, thereby achieving higher-precision global optimization under calibration conditions.

[0050] For example, in the solution process, this invention employs a second-order optimization strategy based on Gauss-Newton iteration and sparse Cholesky decomposition to achieve global pose and geometric consistency. Each keyframe contains 7 degrees of freedom pose variables. For N keyframes, 14×14 sub-blocks are formed and accumulated to a 7N×7N sparse Hessian matrix. Depending on whether the camera intrinsics are known, the optimization residuals are defined in either the ray space or the pixel space: when the camera intrinsics are unknown, the residuals are constrained by ray direction consistency; when the camera intrinsics are known, the residuals are constrained by pixel projection error. To avoid degradation problems in pure rotation scenes, a distance consistency term with a small weight is introduced into the objective function. This optimization is performed for at most 10 iterations after each new keyframe is added and terminates early upon convergence.

[0051] After completing loop closure detection and second-order optimization, Gaussian scene construction is performed. The system projects 3D Gaussian points onto a 2D image plane and uses tile-based rasterization and blending techniques for rapid sorting and fusion, thereby achieving efficient color rendering. Pixels The color is determined by the Gaussian set of all elements that affect that pixel. The decision is made and calculated as follows:

[0052] in, Indicates the pixel position The final rendered color value at that location. Indicates the first The color of a Gaussian in that pixel direction No. An opacity of one Gaussian unit. This indicates the degree to which the pixel has not been completely occluded before the i-th Gaussian. By jointly optimizing parameters such as position, rotation, scale, color, and transparency, the system can achieve fast dynamic rendering while maintaining geometric accuracy.

[0053] Within a local window, all keyframes and their corresponding 3D Gaussian parameters are jointly optimized. The optimization objective is to minimize photometric error, while introducing isotropic constraints to prevent excessive deformation of the Gaussian ellipsoid. The regularization term is defined as follows:

[0054]

[0055] in, This represents the isotropic regularization loss. This represents the number of Gaussians in the set. Indicates the first A Gaussian scaling parameter, For its target scale, This represents the first norm, used to constrain the scale differences of the 3D Gaussian in various directions. This constraint can effectively suppress overstretching of the ellipsoid during the optimization process.

[0056] Taking into account both photometric error and regularization constraints, the objective function for optimizing the local window can be expressed as:

[0057]

[0058] in, Indicates the first The pose transformation matrix of the frame camera from the world coordinate system to the camera coordinate system. Indicates the first Photometric consistency error corresponding to the frame. The weight coefficients for the regularization term are denoted as . By iteratively solving the above objective function, the joint optimization of keyframe pose and Gaussian parameters can be achieved, thereby obtaining a high-precision, continuously differentiable 3D scene representation.

[0059] By leveraging analytical Jacobian matrices and parallel reduction to accelerate the construction of Hessian matrices using CUDA, this invention's backend optimization can achieve efficient global pose and geometric consistency correction without increasing system bottlenecks. Compared to traditional first-order methods, the second-order information introduced in this method significantly improves convergence speed and global optimization quality.

[0060] Furthermore, to achieve real-time rendering and high-precision reconstruction, a scene representation method based on a 3D Gaussian distribution is adopted. This method uses a Gaussian distribution as the smallest geometric unit and achieves joint modeling of geometry and photometric properties through continuously differentiable volume rendering. Each Gaussian distribution is represented by a point... Centered on, its spatial shape and orientation are determined by the covariance matrix. Description. The Gaussian distribution is defined as follows:

[0061]

[0062] in, This represents any position within the 3D scene. To enhance the model's geometric control, the covariance matrix is ​​decomposed into a rotation matrix. With scale matrix Combinations:

[0063]

[0064] In addition, to improve modeling efficiency and reduce redundancy, this invention adopts a sparse hierarchical sampling strategy, retaining only Gaussian points that significantly contribute to the scene, thereby effectively reducing computational load and storage pressure, and ensuring high-quality scene reconstruction.

[0065] Furthermore, to maintain the geometric consistency of the map and avoid keyframe redundancy, this invention employs a keyframe management strategy based on common-view relationships during the mapping process. The system maintains a window for local keyframes. It is used to manage the set of keyframes observing the same spatial region. The selection of new keyframes is based on the viewpoint overlap with the most recent keyframe. When the overlap is lower than a set threshold, the window content is automatically updated to ensure that local constraints are sufficient rather than redundant.

[0066] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0067] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A Gaussian representation SLAM method based on dense matching priors and factor graph constraints, characterized in that: The following steps are included: Step S1: Input the current image and keyframe image, use the pre-trained Vision Transformer model to extract feature points and construct point clouds, output the point map corresponding to the image, and return the matching status of points in the two frames and their respective point cloud information. Step S2: Based on the point-level matching results, high-precision pose estimation optimization of the current frame is obtained through ray constraint optimization. Using the optimized relative pose, the point map of the key frame is spatially transformed and fused to update, a locally consistent point cloud map is constructed, Gaussian scene representation and rasterization rendering are performed, visualization reconstruction and secondary pose correction are completed, and point map data with high-precision pose is generated and transmitted to the backend. Step S3: The backend receives pose and point cloud data, performs loop closure detection, identifies repeated keyframes, introduces global constraints and performs second-order nonlinear global optimization when a loop closure is detected, and establishes local constraints to maintain map continuity and scale consistency when no loop closure is detected. Gaussian rendering is performed through the optimized keyframe map to complete the global dense 3D reconstruction. The backend receives pose optimization results and point cloud data from the frontend, manages the keyframe sequence output by the frontend, and when the number of effective matches between consecutive frames or the number of unique keyframe pixels is lower than the threshold, the system automatically adds new keyframes and establishes bidirectional edges between them and the previous keyframe to add them to the edge set, ensuring the continuity of adjacent keyframes in the time series. Performing loop closure detection includes: querying the database using the encoded features of the current keyframe, and retrieving the most similar previous keyframe. Each candidate keyframe is represented by clustering quantization to encode features, and feature matching and quantization are achieved through L2 distance calculation. When the retrieval score is higher than the set threshold, the candidate keyframe pairs are input to the dual-input Transformer decoder module for dense feature matching. When the number of matched feature points exceeds the threshold, a bidirectional constraint edge is added to the frame pair in the keyframe graph to form a closed-loop constraint. The encoded features of new keyframes are inserted into the inverse file index structure, and the retrieval database is dynamically updated to achieve online updating and response of loop closure detection; Introducing global constraints and performing second-order nonlinear global optimization includes: after loop closure detection, proceeding to backend global optimization; given the current pose estimate of the keyframe and its corresponding standard point map, performing global consistency optimization on the pose and geometry of all keyframes; optimizing based on a global error model with ray constraints when the camera intrinsic parameters are unknown, and optimizing based on a global error model with pixel projection constraints when the camera intrinsic parameters are known; during the global optimization solution process, a second-order optimization strategy based on Gauss-Newton iteration and sparse Cholesky decomposition is adopted to achieve global pose and geometry consistency.

2. The Gaussian representation SLAM method based on dense matching prior and factor graph constraints according to claim 1, characterized in that: The optimization of obtaining high-precision pose estimation of the current frame through ray constraint optimization includes: calling the Sim(3) model based on ray distance constraint or the SE(3) model based on intrinsic parameter constraint to solve for high-precision pose.

3. The Gaussian representation SLAM method based on dense matching priors and factor graph constraints according to claim 2, characterized in that: The optimization of obtaining high-precision pose estimation for the current frame through ray constraint optimization also includes: When the camera intrinsic parameters are known, an optimization model based on ray direction error is used to replace the depth-sensitive point error term. After converting the point map into a unit direction vector form, ray direction difference constraints are applied to each pair of matching points. The system uses keyframes as its core, estimates the relative transformation matrix between the current frame and the previous keyframe, defines the point map estimation of the known previous keyframe, obtains the 3D points corresponding to the current frame through the feature matching module, and solves the pose transformation. After each pose calculation is completed, the point map of the key frame is updated using a weighted average filtering strategy, and observation information from different perspectives is fused to achieve enhanced geometric consistency. After obtaining the optimized camera pose parameters, a 3D Gaussian scene representation is introduced to perform differentiable photometric consistency optimization.

4. The Gaussian representation SLAM method based on dense matching priors and factor graph constraints according to claim 3, characterized in that: In solving the pose transformation process, when the camera intrinsic parameters are unknown, pose estimation is performed using ray distance constraints with consistent scale; when the camera intrinsic parameters are known, pose optimization is performed using ray direction error.

5. The Gaussian representation SLAM method based on dense matching prior and factor graph constraints according to claim 4, characterized in that: Introducing a 3D Gaussian scene representation for differentiable photometric consistency optimization includes: Construct a continuous volume model of the scene to align the rendering result with the real image at the pixel level, and define a photometric loss function; We introduce the Gauss-Newton iteration and iterative reweighted least squares framework to calculate the analytical Jacobian matrix of pose perturbation by ray and distance residuals. By constructing the residual matrix, Jacobian matrix, weight matrix, and relative transformation matrix, we iteratively solve the linear equation and update the pose.

6. The Gaussian representation SLAM method based on dense matching prior and factor graph constraints according to claim 5, characterized in that: The second-order optimization strategies based on Gauss-Newton iteration and sparse Cholesky decomposition include: Each keyframe contains 7 degrees of freedom pose variables. For N keyframes, 14×14 sub-blocks will be formed and accumulated to a sparse Hessian matrix of 7N×7N. Depending on whether the camera intrinsics are known, the optimization residuals are defined in either the ray space or the pixel space: when the camera intrinsics are unknown, the residuals are constrained by the consistency of ray direction; when the camera intrinsics are known, the residuals are constrained by the pixel projection error. A distance consistency term with a small weight is introduced into the objective function. This optimization is performed for at most 10 iterations after each new keyframe is added and terminates early upon convergence. After completing loop closure detection and second-order optimization, Gaussian scene construction is performed, and three-dimensional Gaussian points are projected onto a two-dimensional image plane. Then, tile-based rasterization and blending techniques are used for sorting and fusion to achieve color rendering. Within a local window, all keyframes and their corresponding 3D Gaussian parameters are jointly optimized to obtain a continuously differentiable 3D scene representation. The Hessian matrix is ​​constructed using CUDA through analytical Jacobian matrix and parallel reduction.

7. The Gaussian representation SLAM method based on dense matching prior and factor graph constraints according to claim 6, characterized in that: Implementing color rendering includes: A scene representation method based on 3D Gaussian distribution is adopted, with Gaussian distribution as the smallest geometric unit. The joint modeling of geometry and photometry is achieved through continuously differentiable volume rendering. Each Gaussian is centered at a point, and its spatial shape and orientation are described by the covariance matrix. The covariance matrix is ​​decomposed into a combination of rotation matrix and scale matrix. A sparse hierarchical sampling strategy is adopted to retain only Gaussian points that contribute highly to the scene. A keyframe management strategy based on shared-view relationships is adopted to maintain a window of local keyframes for managing the set of keyframes observing the same spatial region. The selection of new keyframes is based on the viewpoint overlap with the nearest keyframe. When the overlap is lower than a set threshold, the window content is automatically updated.

Images