A 3D GS fusion method for multi-station data acquisition

Abstract

This invention provides a 3DGS fusion method for multi-station data acquisition. It utilizes a handheld device to collect data from the same scene in segments, and performs SLAM mapping on each segment. The separately processed point cloud data are automatically stitched together according to the acquisition order. After successful stitching, new IMUPOS data is recalculated. 3DGS pre-training is performed using the stitched IMUPOS and the original image data, and the individual 3DGS pre-training results are merged. 3DGS training is then performed using the merged 3DGS pre-training results. This invention generates a unified global coordinate system and optimized IMUPOS data after point cloud stitching, providing a consistent initial pose for 3DGS pre-training at each station and reducing local optimization bias. During point cloud stitching, graph optimization and loop closure detection suppress accumulated errors, allowing subsequent 3DGS pre-training to be performed on a more accurate pose basis, significantly reducing reprojection errors.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a 3DGS fusion method for multi-station data acquisition. Background Technology

[0002] 3DGS, as a new 3D reconstruction technology, has been widely applied. Handheld devices are currently the primary method for acquiring 3DGS data. The main advantage of handheld acquisition is its flexibility, but its disadvantages are also significant. For large scenes, a single acquisition cannot fully cover them, requiring multiple acquisitions to cover the entire scene. The data from multiple acquisitions needs to be trained separately to produce 3DGS results, which are then manually merged from multiple 3DGS results of the same scene. This process has obvious drawbacks: firstly, the stitching accuracy of multiple 3DGS results is poor; secondly, noise in each 3DGS image can severely affect the display of adjacent 3DGS images.

[0003] The traditional multi-station 3DGS processing workflow (first SLAM mapping, then separate pre-training, and finally stitching together the 3DGS results) has two main problems: First, the local errors accumulated during SLAM mapping at each station are amplified during stitching, resulting in inconsistencies in the global coordinate system; second, the initial poses used during 3DGS pre-training at each station are not uniform, which can easily lead to pose conflicts and accumulation of reprojection errors after merging. Summary of the Invention

[0004] In view of this, the present invention aims to overcome the shortcomings of the prior art by proposing a 3DGS fusion method for multi-station data acquisition. This method involves first performing laser SLAM mapping on multi-station data acquired in segments within the same scene, then stitching the mapping results together. The stitched point clouds are then used for 3DGS pre-training, and the pre-training results are merged. Finally, the merged data is used for 3DGS training. This significantly improves the fusion accuracy of multi-station data and avoids the display of noise between segmented 3DGS training results.

[0005] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0006] A 3DGS fusion method for multi-station data acquisition includes the following steps:

[0007] Step 1: Use a handheld device to collect raw image data of the same scene in segments, and perform SLAM mapping on the collected raw image data to obtain point cloud data;

[0008] Step 2: Automatically stitch the processed point cloud data according to the acquisition order. After successful stitching, recalculate the new IMUPOS data.

[0009] Step 3: Perform 3DGS pre-training on the newly calculated IMUPOS data and the original image data respectively, and then merge the individual 3DGS pre-training results;

[0010] Step 4: Use the merged 3DGS pre-training results to train 3DGS and output the reconstruction results.

[0011] Furthermore, step 1 includes using a lidar device to scan the surrounding environment by emitting a laser beam, calculating the distance to obstacles based on the laser reflection time or phase difference, generating a three-dimensional point cloud by combining its own rotation angle, recording the position and shape information of objects in the environment, and the SLAM mapping system comparing the point clouds at adjacent time points and using a registration algorithm to calculate the optimal transformation matrix between the point clouds to obtain the relative pose; at the same time, it combines the initial pose estimation of the inertial measurement unit (IMU) and the odometry device to fuse the real-time pose.

[0012] Furthermore, during the mapping process of the SLAM mapping system, a lightweight semantic segmentation network is introduced to identify dynamic objects in real time, and dynamic objects are removed during the point cloud generation stage. The filtering function is as follows:

[0013] ;

[0014] in This is the static point cloud set obtained after filtering. Let i be the i-th point in the original point cloud. This is the original, complete collection of point clouds. This is a semantic classification function that returns the predicted class label of a point.

[0015] By overlaying point clouds at different times according to the estimated pose, a global map is gradually constructed. Loop closure detection identifies previously visited areas, and graph optimization algorithms are used to correct the global pose, ultimately resulting in a consistent and accurate environmental map.

[0016] Furthermore, step 2 specifically includes:

[0017] A point cloud registration algorithm is used to register adjacent point cloud data.

[0018] Based on the pose transformation obtained from point cloud registration, the current point cloud to be stitched is transformed into the coordinate system of the reference point cloud to achieve point cloud stitching.

[0019] After completing the point cloud stitching, pose optimization is performed based on the stitched point cloud map and inertial measurement unit (IMU) data.

[0020] Based on the optimized point cloud pose and inertial measurement unit data, the IMUPOS data is recalculated.

[0021] Furthermore, the point cloud registration algorithm includes: coarse registration using FPFH feature descriptors, followed by fine registration using ICP, with the registration error function being:

[0022]

[0023] in For rotation matrix, It is a translation vector. For the j-th point in the target point cloud, Let i be the i-th point in the source point cloud. For feature distance weights, It is a distance metric function based on FPFH feature descriptors.

[0024] Furthermore, the pose optimization includes: constructing an optimization model that includes point cloud pose constraints and inertial measurement unit (IMU) pose constraints. The point cloud pose constraints are derived from the relative pose relationships between adjacent point clouds obtained during point cloud registration, and the inertial measurement unit (IMU) pose constraints are derived from the acceleration and angular velocity information measured by the inertial measurement unit (IMU). By minimizing the error function in the model, the optimal pose of all point clouds in the world coordinate system is solved.

[0025] During the optimization process, a consistency constraint on semantic feature points is added, and the optimization objective function is:

[0026]

[0027] in The pose constraint error originates from the relative pose relationship of point cloud registration; The IMU constraint error originates from motion constraints measured by the IMU; the semantic consistency constraint error... , This is a camera projection function that projects 3D points onto a 2D image plane. The three-dimensional coordinates of the semantic feature points. is the projection point in the image.

[0028] Furthermore, during data reassembly, a sliding window mechanism is used to pre-integrate recent IMU data to avoid the accumulation of global integration errors. The pre-integration formula is:

[0029] ;

[0030] Time interval The change in position within; Time interval The change in velocity within; Let be the velocity vector at time t; Let be the acceleration vector at time t.

[0031] Furthermore, in step 3, the 3DGS pre-training includes:

[0032] Based on timestamp matching, a one-to-one correspondence is established between the newly calculated IMUPOS data and the original image data, ensuring that each image can be associated with the corresponding initial pose;

[0033] Feature extraction is performed on the original image data to obtain the feature points and their descriptors for each image. Based on the similarity of the feature descriptors, feature matching is performed between adjacent images to obtain matching feature point pairs. Combined with the initial pose of the image, the initial coordinates of the feature points in three-dimensional space are calculated through triangulation to form a three-dimensional point cloud reference for the scene.

[0034] Furthermore, merging the individual 3DGS pre-training results includes:

[0035] Analyze 3DGS pre-trained data from multiple sorties that have been unified to the same coordinate system;

[0036] Fusion of sparse point cloud data from multiple 3DGS pre-training results;

[0037] A pose graph optimization model is constructed, with image pose as nodes and relative pose constraints of adjacent images as edges. The reprojection error of the fused point cloud is introduced as a global constraint.

[0038] Merged and integrated point cloud data.

[0039] Compared with existing technologies, the 3DGS fusion method for multi-station data acquisition described in this invention has the following advantages:

[0040] This invention generates a unified global coordinate system and optimized IMUPOS data after point cloud stitching, providing a consistent initial pose for 3DGS pre-training at each station and reducing local optimization deviations.

[0041] In the point cloud stitching process of this invention, graph optimization and loop closure detection are used to suppress accumulated errors, so that subsequent 3DGS pre-training can be carried out on a more accurate pose basis, significantly reducing reprojection errors. Attached Figure Description

[0042] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0043] Figure 1 This is a schematic diagram of a 3DGS fusion method for multi-station data acquisition according to the present invention;

[0044] Figure 2 This is a schematic diagram of the interface of the multi-station data stitching software of the present invention. Detailed Implementation

[0045] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0046] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0047] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0048] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0049] refer to Figure 1 This invention provides a 3DGS fusion method for multi-station acquired data, comprising the following steps:

[0050] Step 1: Use a handheld device to collect data in segments from the same scene, and perform SLAM mapping on the collected data separately;

[0051] Data collection is conducted in segments at multiple stations, requiring that the end point of the previous station's data collection is the start point of the next station's data collection.

[0052] The collected data are first used to build SLAM maps. After the maps are built, it is necessary to ensure that the point cloud data of each station does not have any abnormalities such as layering or deformation.

[0053] LiDAR scans the surrounding environment by emitting laser beams, calculates the distance to obstacles based on the laser reflection time or phase difference, and generates a three-dimensional point cloud by combining its own rotation angle, thus recording the position and shape information of objects in the environment.

[0054] SLAM mapping systems compare point clouds at adjacent time points and use registration algorithms (such as ICP) to calculate the optimal transformation matrix between point clouds to obtain relative pose (position and attitude). Simultaneously, by combining initial pose estimates from inertial measurement units (IMUs) and odometry devices (wheel speedometers, visual odometry, etc.), a more reliable real-time pose is fused. This invention uses IMU and wheel speedometer fusion as an example; in implementation, other positioning sensors can be fused according to the device configuration.

[0055] Traditional methods for fusing IMU and wheel speedometer often employ simple complementary filtering. This invention introduces an adaptive extended Kalman filter (AEKF) to dynamically adjust the noise covariance. and This suppresses the influence of non-Gaussian noise and improves the accuracy of pose estimation in dynamic environments. The process noise covariance matrix represents the uncertainty in the state prediction process, which mainly comes from the measurement noise of the IMU sensor, model simplification error, etc. The observation noise covariance matrix represents the noise level in the observation data, mainly originating from measurement errors in the wheel speedometer and odometer, data synchronization errors, etc. The state vector is defined as:

[0056]

[0057] in For location, For speed, The pose is represented by quaternions. The prediction step is based on IMU measurements.

[0058]

[0059] in The state prediction value at time k. The state transition function is based on the IMU motion model. This is the control input (IMU measurement) at time k. It represents process noise and follows a Gaussian distribution.

[0060] The observation steps are fused with wheel speed and odometer data:

[0061]

[0062] The observation value at time k (wheel speed meter and odometer data). For the observation function, the state is mapped to the observation space. The noise is observed to follow a Gaussian distribution.

[0063] In the point cloud mapping process, the SLAM mapping system introduces a lightweight semantic segmentation network (such as MobileNet-SSD) to identify dynamic objects (such as pedestrians and vehicles) in real time, and removes dynamic objects during the point cloud generation stage to improve the static consistency of the map. The filtering function is as follows:

[0064] ;

[0065] in This is the static point cloud set obtained after filtering. Let i be the i-th point in the original point cloud. This is the original, complete collection of point clouds. This is a semantic classification function that returns the predicted class label of a point.

[0066] By overlaying point clouds from different times according to their estimated poses, a global map is gradually constructed. However, the pose errors accumulated during the motion process will cause map shifts. Therefore, loop closure detection is needed to identify previously visited areas, and then graph optimization algorithms are used to correct the global pose, ultimately obtaining a consistent and accurate environmental map.

[0067] Step 2: The processed point cloud data are automatically stitched together according to the acquisition order. After successful stitching, new IMUPOS data is recalculated. In this invention, "IMUPOS" refers to the position and attitude information obtained from the data of the Inertial Measurement Unit (IMU), including timestamp, three-dimensional position coordinates (x, y, z) and attitude parameters (quaternions). IMUPOS data provides initial pose estimation for the image and is an important input for subsequent three-dimensional reconstruction and optimization.

[0068] Traditional ICP is prone to getting trapped in local optima. This invention combines FPFH feature descriptors for coarse registration, and then uses ICP for fine registration. The registration error function is:

[0069]

[0070] in Given a rotation matrix, a 3×3 rotation transformation; The translation vector is a 3D translation amount. The j-th point in the target point cloud; Let i be the i-th point in the source point cloud; The feature distance weights are set to 0.3 to 0.5. It is a distance metric function based on FPFH feature descriptors.

[0071] Based on the pose transformation obtained from point cloud registration, the current point cloud to be stitched is transformed into the coordinate system of the reference point cloud, thereby achieving point cloud stitching. Specifically, for each point in the source point cloud... By transforming the formula Transform it to the coordinate system of the target point cloud to obtain new points. All transformed points are merged with the reference point cloud to form a new stitched point cloud. Following the acquisition sequence, the above registration and stitching operations are performed on each frame of the point cloud sequentially, ultimately resulting in a complete stitched map containing all point cloud data. For example... Figure 2 As shown.

[0072] After point cloud stitching is completed, pose optimization is performed based on the stitched point cloud map and IMU data. Using graph optimization and other methods, an optimization model is constructed that includes both point cloud pose constraints and IMU pose constraints. In this model, the point cloud pose constraints are derived from the relative pose relationships between adjacent point clouds obtained during point cloud registration, while the IMU pose constraints are derived from the acceleration and angular velocity information measured by the IMU. By minimizing the error function in the model, the optimal pose of all point clouds in the world coordinate system is solved.

[0073] Based on the traditional pose graph, a consistency constraint on semantic feature points is added to improve optimization stability. The optimization objective function is:

[0074]

[0075] in The pose constraint error originates from the relative pose relationship of point cloud registration; IMU constraint error, originating from motion constraints measured by the IMU; semantic consistency constraint error. Proj() is the camera projection function that projects 3D points onto a 2D image plane. The three-dimensional coordinates of the semantic feature points. is the projection point in the image.

[0076] Based on the optimized point cloud pose and IMU data, the IMUPOS data is recalculated. Using the IMU's motion model and the optimized point cloud pose information, the IMU's position and orientation are updated through methods such as integration.

[0077] During the recalculation process, methods such as Kalman filtering can be used to fuse and filter IMU data, further improving the accuracy and stability of IMUPOS data. Kalman filtering can optimally estimate the state (position, attitude, etc.) of the IMU based on its measured and predicted values, effectively reducing the impact of measurement noise and system errors.

[0078] A sliding window mechanism is used to pre-integrate recent IMU data to avoid the accumulation of global integration errors. The pre-integration formula is:

[0079]

[0080] Time interval The change in position within; Time interval The change in velocity within; Let be the velocity vector at time t; Let be the acceleration vector at time t. Local BA optimization is performed within the window to further improve IMUPOS timing consistency.

[0081] Step 3: Use the calculated IMUPOS and the original image data to perform 3DGS pre-training separately, and then merge the individual 3DGS pre-training results;

[0082] The stitched IMUPOS data provides an initial pose reference for 3DGS pre-training, and the visual features of the original image data are combined for joint optimization to finally obtain the high-precision pose of each photo.

[0083] The resolved IMUPOS data contains sensor position (x, y, z) and pose (such as quaternions or rotation matrices) information corresponding one-to-one with the original image acquisition time, and the global coordinate system has been unified through point cloud stitching. Prepare a sequence of original image data synchronized with the IMUPOS data. The images must cover the complete view of the target scene, and each image must be accompanied by camera intrinsic parameters (focal length, principal point coordinates, etc.).

[0084] Based on timestamp matching, a one-to-one correspondence is established between the solved IMUPOS data and the original image data, ensuring that each image can be associated with its corresponding initial pose (from IMUPOS). The original image data is preprocessed, including removing blurry, overexposed, or underexposed abnormal images, and performing distortion correction on the images (based on camera intrinsics) to improve the stability of subsequent feature extraction.

[0085] Feature extraction is performed on the original image data to obtain feature points and their descriptors for each image. Based on the similarity of the feature descriptors, feature matching is performed between adjacent images to obtain matched feature point pairs. Combined with the initial pose of the image, the initial coordinates of the feature points in 3D space are calculated through triangulation to form a 3D point cloud reference for the scene.

[0086] Dynamically adjust feature extraction parameters for different scene types (e.g., open space, general scene, night scene):

[0087] In open scenes: Use SIFT features and increase the threshold to suppress noise;

[0088] Typical scenario: Use ORB features to balance speed and stability;

[0089] Night mode: First, image enhancement is performed, and then the feature point matching rate under low light conditions is improved.

[0090] Merging individual 3DGS pre-training results includes

[0091] We analyze 3DGS pre-trained data from multiple stations that have been unified to the same coordinate system. Each dataset contains a camera intrinsics file (cameras.txt), an image pose file (images.txt), a sparse point cloud file (points3D.txt), and the corresponding image file. We analyze each dataset to extract key information: camera intrinsics (focal length, distortion coefficients, etc.), image pose (rotation matrix R, translation vector t), and point cloud data (3D coordinates X / Y / Z, color RGB).

[0092] Sparse point cloud data from multi-station 3DGS pre-training results are fused. A clustering algorithm is used to remove duplicate point clouds in overlapping regions, with a fusion threshold set to 0.05m. The fused point cloud has uniform density, avoiding redundancy in overlapping regions.

[0093] A pose graph optimization model is constructed: using image poses as nodes and relative pose constraints of adjacent images as edges, the reprojection error of the fused point cloud is introduced as a global constraint. A graph optimization algorithm is used to minimize the overall error and correct local pose deviations, ensuring that the reprojection error of all images is less than 1 pixel.

[0094]

[0095] in These are the relative pose observations between images i and j; This is a function for calculating relative pose. Let be the covariance matrix of the relative pose observations; For three-dimensional points Projection in an image; These are the two-dimensional feature points actually observed in the image; This is the Huber kernel function, used to suppress the effects of outliers; The confidence weights for point clouds are determined by point cloud density and matching quality.

[0096] The merged point cloud data generates a new points3D.txt file, containing globally unique point IDs, coordinates, colors, and associated image indexes. The optimized image poses are merged to generate a new images.txt file, ensuring the pose data correctly corresponds to the image file paths. The camera intrinsic parameter files cameras.txt are merged, duplicate intrinsic parameter items are removed, and camera ID numbers are standardized.

[0097] This invention employs an intrinsic parameter hash table mechanism to automatically identify and merge identical intrinsic parameters, avoiding duplication. Verification logic ensures that each camera ID is unique and that the intrinsic parameters correspond correctly to the image.

[0098] The intrinsic parameter hash table mechanism works by converting camera intrinsic parameters into unique hash key values, enabling fast lookup and deduplication.

[0099] The hash table operation process includes:

[0100] 1) Initialization: Create an empty hash table with the key being the intrinsic parameter hash value and the value being the list of camera IDs and intrinsic parameter details;

[0101] 2) Insertion of internal parameters:

[0102] a) Calculate the hash value for the new internal parameter;

[0103] b) If the hash value does not exist, create a new entry;

[0104] c) If the hash value exists, compare the detailed parameters (considering floating-point errors).

[0105] 3) Merging process: Internal parameters with the same hash value share the same camera ID to avoid duplication.

[0106] Fault tolerance mechanisms include:

[0107] 1) Floating-point tolerance: Set a reasonable floating-point comparison tolerance (e.g., 0.001);

[0108] 2) Hash collision handling: Hash collisions are handled using a linked list method to ensure data integrity.

[0109] This invention incorporates a verification logic design to ensure the correctness and completeness of internal parameter merging. The verification logic execution flow is as follows:

[0110] Preprocessing stage:

[0111] a) Load all camera intrinsic data;

[0112] b) Create an index for the internal parameter hash table;

[0113] c) Prepare the image samples required for verification;

[0114] Layered inspection and calibration execution:

[0115] a) Level 1: Camera-by-camera basic parameter verification;

[0116] b) Level 2: Photometric consistency verification based on actual images;

[0117] c) Level 3: Overall consistency verification of the multi-camera system;

[0118] Result processing:

[0119] a) Generate a detailed calibration report;

[0120] b) Mark the problematic camera and internal parameters;

[0121] c) Provide repair suggestions.

[0122] Automated repair:

[0123] a) Automatically adjust to a reasonable range for minor parameter deviations;

[0124] b) For serious problems, prompt for manual intervention.

[0125] The advantages of the verification logic of this invention include:

[0126] Comprehensiveness: Covering multiple levels from basic parameters to practical applications.

[0127] Automation: Reduces manual inspection workload and improves processing efficiency

[0128] Fault tolerance: Properly handle floating-point errors and sensor noise.

[0129] Scalability: Facilitates the addition of new verification rules and validation methods.

[0130] Step 4: Use the merged 3DGS pre-training results to train 3DGS.

[0131] Before training with 3DGS, users can select the scene type (open / normal / night scene) and set the zoom factor (default 1.2) through the interface. The system will automatically adjust the training parameters based on the selection.

[0132] In open scenes: the Gaussian kernel size is increased by 20% to avoid model overfitting;

[0133] General scenario: Use default parameters;

[0134] Night mode: Image preprocessing enhances brightness, and the initial brightness value of the Gaussian point is increased;

[0135] Scaling view factor: User adjustable range 1.0~2.0, used to control the distribution density of Gaussian points and the rendering range to adapt to different scene complexities.

[0136] Final output format: The final reconstruction results output by this method include the following files:

[0137] File: Contains optimized 3DGS model point cloud data;

[0138] The cameras.txt file records the intrinsic parameters, extrinsic parameters, and pose information of all cameras;

[0139] The colmap.json file is used to map temporal indexes to image-pose mappings.

[0140] The images.txt file contains a table mapping image paths to poses.

[0141] The points3D.txt file contains the 3D coordinates and color information of sparse point clouds.

[0142] This invention conducts multi-station data collection experiments in an indoor lobby setting (approximately 200㎡) to compare the accuracy of the two methods:

[0143] Method A (traditional training followed by stitching): average reprojection error is 2.8 pixels, point cloud alignment error is 0.15m.

[0144] Method B (this invention first stitches and then trains): the average reprojection error is 1.2 pixels, and the point cloud alignment error is 0.06m.

[0145] Experimental results show that the method of the present invention significantly improves both reprojection error and alignment accuracy, and the merged 3DGS model is visually more coherent and has fewer noise points.

[0146] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. 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 3D GS fusion method of multi-station acquisition data, characterized in that: Includes the following steps: Step 1: Use a handheld device to collect raw LiDAR point cloud data and raw image data in the same scene in segments, and perform SLAM mapping on the collected raw LiDAR point cloud data to obtain point cloud data. Step 2: Automatically stitch the processed point cloud data according to the acquisition order. After successful stitching, recalculate the new IMUPOS data; specifically including: A point cloud registration algorithm is used to register adjacent point cloud data. Based on the pose transformation obtained from point cloud registration, the current point cloud to be stitched is transformed into the coordinate system of the reference point cloud to achieve point cloud stitching. After completing the point cloud stitching, pose optimization is performed based on the stitched point cloud map and inertial measurement unit (IMU) data. Based on the optimized point cloud pose and inertial measurement unit data, the IMUPOS data is recalculated; Step 3: Perform 3DGS pre-training on the newly calculated IMUPOS data and the original image data respectively, and then merge the individual 3DGS pre-training results; 3DGS pre-training includes: Based on timestamp matching, a one-to-one correspondence is established between the newly calculated IMUPOS data and the original image data, ensuring that each image can be associated with the corresponding initial pose; Feature extraction is performed on the original image data to obtain feature points and their descriptors for each image. Based on the similarity of the feature descriptors, feature matching is performed between images acquired in adjacent order to obtain matched feature point pairs. Combined with the initial pose of the image, the initial coordinates of the feature points in 3D space are calculated through triangulation to form a 3D point cloud reference for the scene. Merging the individual 3DGS pre-training results includes: Analyze 3DGS pre-trained data from multiple stations that have been unified to the same coordinate system; Fusion of sparse point cloud data from multi-station 3DGS pre-training results; A pose graph optimization model is constructed, with image pose as nodes and relative pose constraints of adjacent images as edges. The reprojection error of the fused point cloud is introduced as a global constraint. Merged and integrated point cloud data; Step 4: Use the merged 3DGS pre-training results to train 3DGS and output the reconstruction results.

2. The 3DGS fusion method for multi-station data acquisition according to claim 1, characterized in that: Step 1 includes using a lidar device to scan the surrounding environment by emitting a laser beam, calculating the distance to obstacles based on the laser reflection time or phase difference, generating a three-dimensional point cloud by combining its own rotation angle, and recording the position and shape information of objects in the environment. The SLAM mapping system compares the point clouds at adjacent time points and uses a registration algorithm to calculate the optimal transformation matrix between the point clouds to obtain the relative pose. At the same time, it combines the initial pose estimation of the inertial measurement unit (IMU) and the odometry device to fuse the real-time pose.

3. The 3DGS fusion method for multi-station data acquisition according to claim 2, characterized in that: In the SLAM mapping system, a lightweight semantic segmentation network is introduced to identify dynamic objects in real time and remove them during the point cloud generation stage: for each point in the original point cloud, the semantic segmentation network determines whether it belongs to a preset dynamic object class. If it does, it is removed, and only static point clouds are retained. The point clouds at different times are superimposed according to the estimated pose to gradually build a global map. The loop closure detection identifies the areas that have been visited, and then the graph optimization algorithm is used to correct the global pose to obtain the environment map.

4. The 3DGS fusion method for multi-station data acquisition according to claim 1, characterized in that: The point cloud registration algorithm includes: first, coarse registration using FPFH feature descriptors, and then fine registration using ICP. The registration error function is: ； in For rotation matrix, It is a translation vector. For the j-th point in the target point cloud, Let i be the i-th point in the source point cloud; For feature distance weights, It is a distance metric function based on FPFH feature descriptors.

5. The 3DGS fusion method for multi-station data acquisition according to claim 1, characterized in that: The pose optimization includes: constructing an optimization model that includes point cloud pose constraints and inertial measurement unit (IMU) pose constraints. The point cloud pose constraints are derived from the relative pose relationships between adjacent point clouds obtained during point cloud registration, and the inertial measurement unit (IMU) pose constraints are derived from the acceleration and angular velocity information measured by the inertial measurement unit (IMU). By minimizing the error function in the model, the optimal pose of all point clouds in the world coordinate system is solved. During the optimization process, a consistency constraint on semantic feature points is added, and the optimization objective function is: ； in The pose constraint error originates from the relative pose relationship of point cloud registration; The IMU constraint error originates from motion constraints measured by the IMU; the semantic consistency constraint error... , This is a camera projection function that projects 3D points onto a 2D image plane. The three-dimensional coordinates of the semantic feature points. is the projection point in the image.

6. The 3DGS fusion method for multi-station data acquisition according to claim 1, characterized in that: When resolving IMUPOS data, a sliding window mechanism is used to pre-integrate recent IMU data to avoid the accumulation of global integration errors. The pre-integration formula is as follows: ； Time interval The change in position within; Time interval The change in velocity within; Let be the velocity vector at time t; Let be the acceleration vector at time t.

Images