Two-dimensional texture mapping method based on multiple subgraphs and storage medium
By leveraging ORB features and a multi-subgraph mechanism based on quadtree distribution, combined with bag-of-words model and Ceres Solver optimization, the efficiency and robustness issues of 2D SLAM ground texture mapping are addressed, achieving efficient and low-cost 2D mapping.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-11
- Publication Date
- 2026-03-24
AI Technical Summary
Existing technologies for 2D SLAM ground texture mapping suffer from problems such as low mapping efficiency in large scenes, high computational complexity, high implementation costs, and low robustness to low-texture scenes.
Feature point extraction is performed using ORB features and a quadtree uniform distribution. The stability of feature matching is improved by Hamming distance and distance ratio testing. A multi-subgraph mechanism is constructed, and a bag-of-words model is used for loop closure detection and global optimization to reduce computational complexity. Finally, Ceres Solver is used for nonlinear optimization.
It significantly improves robustness in low-texture scenes, reduces computational complexity, increases mapping speed and efficiency, lowers implementation costs, and maintains map accuracy and stability.
Smart Images
Figure CN121725104A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of two-dimensional mapping technology, and in particular to a two-dimensional texture mapping method and storage medium based on multiple subgraphs. Background Technology
[0002] In the field of 2D SLAM (Simultaneous Localization and Mapping), ground texture mapping technology is crucial for applications such as AGV navigation and warehouse robots. Existing technologies primarily employ LiDAR for 2D environment mapping, employing methods such as: 1. Enhancing the accuracy and robustness of 2D mapping by utilizing ground texture features; 2. Improving the efficiency of large-scene mapping through adaptive sub-graph segmentation; and 3. Solving the problem of stable matching of texture features in changing environments. However, these methods have the following limitations:
[0003] 1. Global optimization is inefficient when building large-scale maps. The computational complexity increases exponentially with the map size, resulting in high implementation costs. Auxiliary mapping tools such as crosshairs need to be laid out before map building.
[0004] 2. Existing subgraph techniques (such as Cartographer) mainly process point cloud data, lack specific optimization mechanisms for ground texture features, and have low robustness to low-texture scenes. Summary of the Invention
[0005] The purpose of this application is to overcome the problems of high implementation cost and low robustness to low-texture scenes in the prior art of texture SLAM, and to provide a two-dimensional texture mapping method and storage medium based on multiple subgraphs.
[0006] Firstly, a method for constructing two-dimensional textures based on multiple subgraphs is provided, including:
[0007] Feature points are extracted using ORB features and a quadtree uniform distribution.
[0008] Hamming distance of descriptors is used for feature matching, and the stability of feature matching is improved by distance ratio testing;
[0009] When the pose transformation exceeds a certain threshold, keyframes are inserted into the texture feature sub-graph. When the localization is accidentally lost during map construction or the keyframes in the sub-graph reach a preset threshold, a new texture feature sub-graph is constructed. The texture feature sub-graph is used as a map unit, and the feature matching results are stored. The feature matching results include a set of keyframes, a feature description subset, and the sub-graph pose.
[0010] The feature matching results are sent to the backend for loop closure detection and global optimization.
[0011] In some possible implementations, feature point extraction is performed using ORB features and a uniform quadtree distribution, including:
[0012] Feature points are extracted using the ORB algorithm:
[0013]
[0014] in, It is the set of all ORB feature points, (x i y i ) represents the coordinates of the feature point, θ i Main direction, d i It is a 256-bit binary descriptor;
[0015] Achieving uniform distribution of feature points using a quadtree:
[0016]
[0017] Where, N max_per_cell The maximum number of points in each cell, N total d represents the total number of feature points, and d represents the depth of the quadtree.
[0018] In some possible implementations, the Hamming distance is calculated using the following formula:
[0019]
[0020] Where, d Hamming Let d be the Hamming distance function. i and d j For two ORB descriptors, ⊕ is the XOR operation, and popcount() is the bit counting function;
[0021] The formula for the distance ratio test is:
[0022]
[0023] Where, d nearest For the closest distance, d second This is the second closest distance.
[0024] In some possible implementations, the feature matching results are stored, including:
[0025]
[0026] Keyframes i Features is a set of keyframes. i For a subset of feature descriptions, Pose i The pose of the subgraph.
[0027] In some possible implementations, the loop closure detection includes:
[0028] Insert the feature matching results into the pose graph:
[0029] G = (V, E)
[0030] Where, V={v i |v i For keyframe nodes}, E={ei j |e ij {For pose constraint edges};
[0031] Keyframes and constraints with other subgraphs are constructed using the bag-of-words model:
[0032] Building a visual dictionary using the DBoW3 bag-of-words model:
[0033]
[0034] Vocabulary is a visual dictionary containing a set of K visual words, w K Let K be the Kth visual word, where K is the size of the visual dictionary;
[0035] Calculate the bag-of-words vectors for keyframes:
[0036]
[0037] Among them, v bow For the bag-of-words vector representation of an image, tf i For the word w i Frequency;
[0038] Calculate the similarity S between keyframes ij :
[0039]
[0040] Among them, v i Let v be the bag-of-words vector of keyframe i. j Let be the bag-of-words vector of keyframe j.
[0041] In some possible implementations, the global graph optimization includes:
[0042] Global graph optimization is performed by constructing constraints between keyframes and subgraphs, as well as constraints on other subgraphs.
[0043]
[0044] Among them, the IMU constraint is:
[0045]
[0046] Cycle constraint:
[0047]
[0048] Subgraph constraints:
[0049]
[0050] Among them, E total It is the sum of all residuals, and is the final optimization objective function, E. odometry E loop and E submap Different residual functions are obtained for different sensors to minimize the error, T i and T j To optimize variables, , and These are the observations obtained through IMU, loop closure, and subgraph matching, respectively. ρ(·) represents the square of the Frobenius norm, and ρ(·) represents the kernel function used to control outliers.
[0051] Handling outliers using robust kernel functions:
[0052]
[0053] in, Here, r is the loss function, r is the difference between the observed value and the model prediction, and δ is a threshold parameter used to control when to switch from squared loss to linear loss.
[0054] Nonlinear optimization using Ceres Solver:
[0055]
[0056] in, To obtain the objective function to be optimized, nonlinear optimization theory is used. i and T j The optimal estimate.
[0057] Secondly, a two-dimensional texture mapping method based on multiple subgraphs is provided, including:
[0058] The feature matching results from the front end are obtained. In the front end, feature subgraphs are constructed as map units by uniformly distributing ORB features and quadtrees, and the feature matching results are stored. The feature matching results include a set of keyframes, a set of feature descriptions, and the pose of the subgraphs.
[0059] The feature matching results are inserted into the pose graph, and keyframes and other subgraph constraints are constructed using the bag-of-words model.
[0060] Global graph optimization is performed by constructing constraints between keyframes and subgraphs, as well as other subgraph constraints. A robust kernel function is used to handle outliers, and Ceres Solver is used for nonlinear optimization.
[0061] In some possible implementations, the feature matching results are inserted into the pose graph, and keyframes and constraints on other subgraphs are constructed using a bag-of-words model, specifically including:
[0062] Insert the feature matching results into the pose graph:
[0063] G = (V, E)
[0064] Where, V={v i |v i For keyframe nodes}, E={ei j |e ij {For pose constraint edges};
[0065] Keyframes and constraints with other subgraphs are constructed using the bag-of-words model:
[0066] Building a visual dictionary using the DBoW3 bag-of-words model:
[0067]
[0068] Vocabulary is a visual dictionary containing a set of K visual words, w K Let K be the Kth visual word, where K is the size of the visual dictionary;
[0069] Calculate the bag-of-words vectors for keyframes:
[0070]
[0071] Among them, v bow For the bag-of-words vector representation of an image, tf i For the word w i Frequency;
[0072] Calculate the similarity S between keyframes ij :
[0073]
[0074] Among them, v i Let v be the bag-of-words vector of keyframe i. j Let be the bag-of-words vector of keyframe j.
[0075] In some possible implementations, global graph optimization is performed by constructing constraints between keyframes and subgraphs, as well as other subgraph constraints. Robust kernel functions are used to handle outliers, and Ceres Solver is used for nonlinear optimization. Specifically, this includes:
[0076] Global graph optimization is performed by constructing constraints between keyframes and subgraphs, as well as constraints on other subgraphs.
[0077]
[0078] Among them, the IMU constraint is:
[0079]
[0080] Cycle constraint:
[0081]
[0082] Subgraph constraints:
[0083]
[0084] Among them, E total It is the sum of all residuals, and is the final optimization objective function, E. odometry E loop and E submap Different residual functions are obtained for different sensors to minimize the error, T i and T j To optimize variables, , and These are the observations obtained through IMU, loop closure, and subgraph matching, respectively. ρ(·) represents the square of the Frobenius norm, and ρ(·) represents the kernel function used to control outliers.
[0085] Handling outliers using robust kernel functions:
[0086]
[0087] in, Here, r is the loss function, r is the difference between the observed value and the model prediction, and δ is a threshold parameter used to control when to switch from squared loss to linear loss.
[0088] Nonlinear optimization using Ceres Solver:
[0089]
[0090] in, To obtain the objective function to be optimized, nonlinear optimization theory is used. i and T j The optimal estimate.
[0091] Thirdly, a computer-readable storage medium is provided for program code executed by a device, the program code including steps for performing a method as described in any of the implementations of the first and second aspects above.
[0092] This application has the following beneficial effects: It can still work stably in low-texture scenes by using ORB features and quadtree distribution, which significantly improves the robustness in low-texture scenes. It reduces the computational complexity through the multi-subgraph mechanism, and improves the mapping speed by more than 40%, which greatly improves the efficiency of 2D mapping. In addition, no auxiliary markers need to be laid in the 2D mapping process of this application, thereby reducing the implementation cost by more than 80%, which greatly reduces the cost of 2D mapping. Furthermore, this application adopts bag-of-words model loop closure detection to ensure the accuracy of the map, making the positioning error less than 5cm and maintaining stable map accuracy. Attached Figure Description
[0093] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.
[0094] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0095] Figure 1 This is a flowchart of the two-dimensional texture mapping method based on multiple subgraphs according to Embodiment 1 of this application;
[0096] Figure 2 This is a flowchart of a two-dimensional texture mapping method based on multiple subgraphs according to Embodiment 2 of this application. Detailed Implementation
[0097] 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.
[0098] Example 1
[0099] like Figure 1 As shown, Embodiment 1 of this application relates to a two-dimensional texture mapping method based on multiple subgraphs, with the front end as the execution entity. The two-dimensional texture mapping method includes:
[0100] S101. Feature point matching at the front end: Feature points are extracted using ORB features and a quadtree with uniform distribution.
[0101] Because the ground texture has a small field of view, feature points tend to cluster together. Therefore, we use ORB (Oriented Fast and Rotated BRIEF) feature points + quadtree to increase the uniformity of feature points.
[0102] Feature points are extracted using the ORB algorithm:
[0103]
[0104] in, It is the set of all ORB feature points, (x i y i ) represents the coordinates of the feature point, θ i Main direction, d i It is a 256-bit binary descriptor;
[0105] Achieving uniform distribution of feature points using a quadtree:
[0106]
[0107] Where, N max_per_cell The maximum number of points in each cell, where Cell refers to a rectangular image region formed after recursively partitioning the quadtree, N total d represents the total number of feature points, and d represents the depth of the quadtree.
[0108] S102. Front-end feature matching: Hamming distance of descriptors is used for feature matching, and distance ratio test is used to improve the stability of feature matching.
[0109] Calculate descriptor similarity using Hamming distance:
[0110]
[0111] Where, d Hamming Let d be the Hamming distance function. i and d j For two ORB descriptors, ⊕ is the XOR operation, and popcount() is the bit counting function;
[0112] The formula for the distance ratio test is:
[0113]
[0114] Where, d nearest For the closest distance, d second This is the second closest distance.
[0115] S103. Front-end performs sub-graph construction: When the pose transformation is greater than a certain threshold, keyframes are inserted into the texture feature sub-graph. When the localization is accidentally lost during map construction or the keyframes in the sub-graph reach a preset threshold, a new texture feature sub-graph is constructed. The texture feature sub-graph is used as a map unit and the feature matching result is stored. The feature matching result includes a set of keyframes, a feature description subset, and the sub-graph pose.
[0116] The feature matching results are as follows:
[0117]
[0118] Keyframes i Features is a set of keyframes. i For a subset of feature descriptions, Pose i The pose of the subgraph.
[0119] S104. The feature matching results are sent to the backend for loop closure detection and global optimization.
[0120] Backend loopback detection includes:
[0121] 1) Insert the feature matching results into the pose graph:
[0122] G = (V, E)
[0123] Where, V={v i |v i For keyframe nodes}, E={ei j |e ij {For pose constraint edges};
[0124] 2) Construct keyframes and constraints with other subgraphs using the bag-of-words model:
[0125] Building a visual dictionary using the DBoW3 bag-of-words model:
[0126]
[0127] Vocabulary is a visual dictionary containing a set of K visual words, w K Let K be the Kth visual word, where K is the size of the visual dictionary;
[0128] Calculate the bag-of-words vectors for keyframes:
[0129]
[0130] Among them, v bow For the bag-of-words vector representation of an image, tf i For the word w i Frequency;
[0131] Calculate the similarity S between keyframes ij :
[0132]
[0133] Among them, v i Let v be the bag-of-words vector of keyframe i. j Let be the bag-of-words vector of keyframe j.
[0134] Global graph optimization includes:
[0135] The backend constructs constraints between keyframes and subgraphs, as well as other subgraph constraints, and performs global graph optimization:
[0136]
[0137] Among them, the IMU constraint is:
[0138]
[0139] Cycle constraint:
[0140]
[0141] Subgraph constraints:
[0142]
[0143] Among them, E total It is the sum of all residuals, and is the final optimization objective function, E. odometry E loop and E submap Different residual functions are obtained for different sensors to minimize the error, T i and T j To optimize variables, , and These are the observations obtained through IMU, loop closure, and subgraph matching, respectively. ρ(·) represents the square of the Frobenius norm, and ρ(·) represents the kernel function used to control outliers.
[0144] Handling outliers using robust kernel functions:
[0145]
[0146] in, Here, r is the loss function, r is the difference between the observed value and the model prediction, and δ is a threshold parameter used to control when to switch from squared loss to linear loss.
[0147] Nonlinear optimization using Ceres Solver:
[0148]
[0149] in, To obtain the objective function to be optimized, nonlinear optimization theory is used. i and T j The optimal estimate.
[0150] In this embodiment, ORB features and quadtree distribution can still work stably in low-texture scenes, significantly improving robustness in low-texture scenes. The multi-subgraph mechanism reduces computational complexity and improves mapping speed by more than 40%, greatly improving the efficiency of 2D mapping. Furthermore, no auxiliary markers need to be laid during the 2D mapping process, thereby reducing implementation costs by more than 80%, significantly reducing the cost of 2D mapping. In addition, bag-of-words model loop closure detection is used in this application to ensure map accuracy, making the positioning error less than 5cm and maintaining stable map accuracy.
[0151] Example 2
[0152] like Figure 2 As shown, Embodiment 2 of this application relates to a two-dimensional texture mapping method based on multiple subgraphs, with a backend as the execution entity. The two-dimensional texture mapping method includes:
[0153] S201. The backend obtains the feature matching results from the frontend. In the frontend, feature subgraphs are constructed as map units by uniformly distributing ORB features and quadtrees, and the feature matching results are stored. The feature matching results include a set of keyframes, a set of feature descriptions, and the pose of the subgraphs.
[0154] In the front end, due to the small field of view of ground texture, feature points tend to cluster together. The uniformity of feature points is increased by using ORB (Oriented Fast and Rotated BRIEF) feature points + quadtree.
[0155] Feature points are extracted using the ORB algorithm:
[0156]
[0157] in, It is the set of all ORB feature points, (x i y i ) represents the coordinates of the feature point, θ i Main direction, d i It is a 256-bit binary descriptor;
[0158] Achieving uniform distribution of feature points using a quadtree:
[0159]
[0160] Where, N max_per_cell The maximum number of points in each cell, where Cell refers to a rectangular image region formed after recursively partitioning the quadtree, N total d represents the total number of feature points, and d represents the depth of the quadtree.
[0161] Feature matching: Hamming distance of descriptors is used for feature matching, and distance ratio test is used to improve the stability of feature matching.
[0162] Calculate descriptor similarity using Hamming distance:
[0163]
[0164] Where, d Hamming Let d be the Hamming distance function. i and d j For two ORB descriptors, ⊕ is the XOR operation, and popcount() is the bit counting function;
[0165] The formula for the distance ratio test is:
[0166]
[0167] Where, d nearest For the closest distance, d second This is the second closest distance.
[0168] Subgraph construction: When the pose transformation is greater than a certain threshold, keyframes are inserted into the texture feature subgraph. When the localization is accidentally lost during map construction or the keyframes in the subgraph reach a preset threshold, a new texture feature subgraph is constructed. The texture feature subgraph is used as a map unit and the feature matching results are stored. The feature matching results include a set of keyframes, a feature description subset, and the subgraph pose.
[0169] The feature matching results are as follows:
[0170]
[0171] Keyframes i Features is a set of keyframes. i For a subset of feature descriptions, Pose i The pose of the subgraph.
[0172] S202. Backend performs loop closure detection: The feature matching results are inserted into the pose graph, and keyframes and other subgraph constraints are constructed using the bag-of-words model.
[0173] Specifically, the feature matching results are inserted into the pose graph:
[0174] G = (V, E)
[0175] Where, V={v i |v i For keyframe nodes}, E={ei j |e ij {For pose constraint edges};
[0176] 2) Construct keyframes and constraints with other subgraphs using the bag-of-words model:
[0177] Building a visual dictionary using the DBoW3 bag-of-words model:
[0178]
[0179] Vocabulary is a visual dictionary containing a set of K visual words, w K Let K be the Kth visual word, where K is the size of the visual dictionary;
[0180] Calculate the bag-of-words vectors for keyframes:
[0181]
[0182] Among them, v bow For the bag-of-words vector representation of an image, tf i For the word w i Frequency;
[0183] Calculate the similarity S between keyframes ij :
[0184]
[0185] Among them, v i Let v be the bag-of-words vector of keyframe i. j Let be the bag-of-words vector of keyframe j.
[0186] S203, the backend constructs constraints between keyframes and subgraphs, as well as other subgraph constraints, to perform global graph optimization, uses a robust kernel function to handle outliers, and uses Ceres Solver for nonlinear optimization.
[0187] Global graph optimization is performed by constructing constraints between keyframes and subgraphs, as well as constraints on other subgraphs.
[0188]
[0189] Among them, the IMU constraint is:
[0190]
[0191] Cycle constraint:
[0192]
[0193] Subgraph constraints:
[0194]
[0195] Among them, E total It is the sum of all residuals, and is the final optimization objective function, E. odometry E loop and E submap Different residual functions are obtained for different sensors to minimize the error, T i and T j To optimize variables, , and These are the observations obtained through IMU, loop closure, and subgraph matching, respectively. ρ(·) represents the square of the Frobenius norm, and ρ(·) represents the kernel function used to control outliers.
[0196] Handling outliers using robust kernel functions:
[0197]
[0198] in, Here, r is the loss function, r is the difference between the observed value and the model prediction, and δ is a threshold parameter used to control when to switch from squared loss to linear loss.
[0199] Nonlinear optimization using Ceres Solver:
[0200]
[0201] in, To obtain the objective function to be optimized, nonlinear optimization theory is used. i and T j The optimal estimate.
[0202] In this embodiment, ORB features and quadtree distribution can still work stably in low-texture scenes, significantly improving robustness in low-texture scenes. The multi-subgraph mechanism reduces computational complexity and improves mapping speed by more than 40%, greatly improving the efficiency of 2D mapping. Furthermore, no auxiliary markers need to be laid during the 2D mapping process, thereby reducing implementation costs by more than 80%, significantly reducing the cost of 2D mapping. In addition, bag-of-words model loop closure detection is used in this application to ensure map accuracy, making the positioning error less than 5cm and maintaining stable map accuracy.
[0203] It should be noted that other specific implementations of the two-dimensional texture mapping device based on multiple subgraphs in this embodiment can be found in the specific implementations of the two-dimensional texture mapping method based on multiple subgraphs described above. To avoid redundancy, they will not be repeated here.
[0204] Example 3
[0205] This application relates to a computer-readable storage medium in Embodiment 3, which is used for program code executed by a device. The program code includes steps for performing the method in any of the implementations of Embodiments 1 and 2 of this application.
[0206] The computer-readable storage medium may be a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM); the computer-readable storage medium may store program code, and when the program stored in the computer-readable storage medium is executed by a processor, the processor is used to perform the steps of the method in any of the implementations of Embodiment 1 of this application.
[0207] The above are merely preferred embodiments of this application; however, the scope of protection of this application is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in this application, based on the technical solution and its improved concept, should be covered within the scope of protection of this application.
Claims
1. A two-dimensional texture mapping method based on multiple subgraphs, characterized in that, include: Feature points are extracted using ORB features and a quadtree uniform distribution. Hamming distance of descriptors is used for feature matching, and the stability of feature matching is improved by distance ratio testing; When the pose transformation exceeds a certain threshold, keyframes are inserted into the texture feature sub-graph. When the localization is accidentally lost during map construction or the keyframes in the sub-graph reach a preset threshold, a new texture feature sub-graph is constructed. The texture feature sub-graph is used as a map unit, and the feature matching results are stored. The feature matching results include a set of keyframes, a feature description subset, and the sub-graph pose. The feature matching results are sent to the backend for loop closure detection and global optimization.
2. The two-dimensional texture mapping method based on multiple subgraphs according to claim 1, characterized in that, Feature point extraction is performed using ORB features and a quadtree uniform distribution, including: Feature points are extracted using the ORB algorithm: ; in, It is the set of all ORB feature points, (x i y i ) represents the coordinates of the feature point, θ i Main direction, d i It is a 256-bit binary descriptor; Achieving uniform distribution of feature points using a quadtree: ; Where, N max_per_cell The maximum number of points in each cell, N total d represents the total number of feature points, and d represents the depth of the quadtree.
3. The two-dimensional texture mapping method based on multiple subgraphs according to claim 2, characterized in that, The formula for calculating the Hamming distance is: ; Where, d Hamming Let d be the Hamming distance function. i and d j For two ORB descriptors, ⊕ is the XOR operation, and popcount() is the bit counting function; The formula for the distance ratio test is: ; Where, d nearest For the closest distance, d second This is the second closest distance.
4. The two-dimensional texture mapping method based on multiple subgraphs according to claim 2, characterized in that, Store the feature matching results, including: ; Keyframes i Features is a set of keyframes. i For a subset of feature descriptions, Pose i The pose of the subgraph.
5. The two-dimensional texture mapping method based on multiple subgraphs according to claim 4, characterized in that, The loop closure detection includes: Insert the feature matching results into the pose graph: G = (V, E); Where, V={v i |v i For keyframe nodes}, E={ei j |e ij {For pose constraint edges}; Keyframes and constraints with other subgraphs are constructed using the bag-of-words model: Building a visual dictionary using the DBoW3 bag-of-words model: ; Vocabulary is a visual dictionary containing a set of K visual words, w K Let K be the Kth visual word, where K is the size of the visual dictionary; Calculate the bag-of-words vectors for keyframes: ; Among them, v bow For the bag-of-words vector representation of an image, tf i For the word w i Frequency; Calculate the similarity S between keyframes ij : ; Among them, v i Let v be the bag-of-words vector of keyframe i. j Let be the bag-of-words vector of keyframe j.
6. The two-dimensional texture mapping method based on multiple subgraphs according to claim 1, characterized in that, The global graph optimization includes: Global graph optimization is performed by constructing constraints between keyframes and subgraphs, as well as constraints on other subgraphs. ; Among them, the IMU constraint is: ; Cycle constraint: ; Subgraph constraints: ; Among them, E total It is the sum of all residuals, and is the final optimization objective function, E. odometry E loop and E submap Different residual functions are obtained for different sensors to minimize the error, T i and T j To optimize variables, , and These are the observations obtained through IMU, loop closure, and subgraph matching, respectively. ρ(·) represents the square of the Frobenius norm, and ρ(·) represents the kernel function used to control outliers. Handling outliers using robust kernel functions: ; in, Here, r is the loss function, r is the difference between the observed value and the model prediction, and δ is a threshold parameter used to control when to switch from squared loss to linear loss. Nonlinear optimization using Ceres Solver: ; in, To obtain the objective function to be optimized, nonlinear optimization theory is used. i and T j The optimal estimate.
7. A two-dimensional texture mapping method based on multiple subgraphs, characterized in that, include: The feature matching results from the front end are obtained. In the front end, feature subgraphs are constructed as map units by uniformly distributing ORB features and quadtrees, and the feature matching results are stored. The feature matching results include a set of keyframes, a set of feature descriptions, and the pose of the subgraphs. The feature matching results are inserted into the pose graph, and keyframes and other subgraph constraints are constructed using the bag-of-words model. Global graph optimization is performed by constructing constraints between keyframes and subgraphs, as well as other subgraph constraints. A robust kernel function is used to handle outliers, and Ceres Solver is used for nonlinear optimization.
8. The two-dimensional texture mapping method based on multiple subgraphs according to claim 7, characterized in that, The feature matching results are inserted into the pose graph, and keyframes and constraints between the keyframes and other subgraphs are constructed using a bag-of-words model. Specifically, this includes: Insert the feature matching results into the pose graph: G = (V, E); Where, V={v i |v i For keyframe nodes}, E={ei j |e ij {For pose constraint edges}; Keyframes and constraints with other subgraphs are constructed using the bag-of-words model: Building a visual dictionary using the DBoW3 bag-of-words model: ; Vocabulary is a visual dictionary containing a set of K visual words, w K Let K be the Kth visual word, where K is the size of the visual dictionary; Calculate the bag-of-words vectors for keyframes: ; Among them, v bow For the bag-of-words vector representation of an image, tf i For the word w i Frequency; Calculate the similarity S between keyframes ij : ; Among them, v i Let v be the bag-of-words vector of keyframe i. j Let be the bag-of-words vector of keyframe j.
9. The two-dimensional texture mapping method based on multiple subgraphs according to claim 7, characterized in that, Global graph optimization is performed by constructing constraints between keyframes and subgraphs, as well as constraints on other subgraphs. A robust kernel function is used to handle outliers, and Ceres Solver is used for nonlinear optimization. Specifically, this includes: Global graph optimization is performed by constructing constraints between keyframes and subgraphs, as well as constraints on other subgraphs. ; Among them, the IMU constraint is: ; Cycle constraint: ; Subgraph constraints: ; Among them, E total It is the sum of all residuals, and is the final optimization objective function, E. odometry E loop and E submap Different residual functions are obtained for different sensors to minimize the error, T i and T j To optimize variables, , and These are the observations obtained through IMU, loop closure, and subgraph matching, respectively. ρ(·) represents the square of the Frobenius norm, and ρ(·) represents the kernel function used to control outliers. Handling outliers using robust kernel functions: ; in, Here, r is the loss function, r is the difference between the observed value and the model prediction, and δ is a threshold parameter used to control when to switch from squared loss to linear loss. Nonlinear optimization using Ceres Solver: ; in, To obtain the objective function to be optimized, nonlinear optimization theory is used. i and T j The optimal estimate.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is for program code executed by the device, the program code including steps for performing the method as described in any one of claims 1-9.