SLAM rear-end pose optimization method based on graph neural network

Abstract

The invention relates to the technical field of simultaneous localization and map construction, and discloses an SLAM (simultaneous localization and mapping) back-end pose optimization method based on a graph neural network, which comprises the following steps of: extracting key feature points of image streams continuously input into an SLAM system, and recording feature descriptors; projecting the key feature points to a three-dimensional space, obtaining corresponding points of left and right camera images through stereo matching, and triangularizing the corresponding points to obtain three-dimensional coordinates of the key feature points; key frames are selected, key feature points and feature descriptors in the key frames are recorded, and whether the number of the key frames is larger than a preset threshold value or not is judged; through key feature point matching, determining a common-view relationship between the key frames, obtaining a graph topological structure of the key frames, and inputting the graph topological structure into a graph neural network for pose optimization; sorting the key frames in the key frame database according to the size of the timestamps in a positive sequence, and keeping the rear half part of the key frames in the key frames sorted in the positive sequence for the next round of data collection and optimization; according to the method, the key frame pose optimization precision is improved, and the performance of the SLAM system is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of simultaneous positioning and map construction, and in particular to a SLAM back-end posture optimization method based on a graph neural network. Background Art

[0002] Positioning and navigation in the field of mobile robots and autonomous driving mainly rely on simultaneous localization and mapping (SLAM) technology. By equipping specific sensors, autonomous unmanned mobile devices can use the SLAM system to build a global map of the unknown environment and achieve accurate positioning and navigation. SLAM technology has important applications in unmanned driving, drones and other fields. In SLAM, finding the correct representation of the corresponding environment and estimating the robot's pose trajectory is important for both positioning and mapping. With the development of science and technology, various SLAM technologies have gradually matured. Pose Graph Optimization (PGO) is an important step in the backend of the SLAM system. However, the existing pose optimization is usually modeled as a nonlinear least squares problem. When the graph is large, the entire pose optimization process will take up a lot of computing resources; at the same time, due to the presence of noise and outliers when the sensor collects environmental data, traditional optimization methods are easily affected by erroneous data and have poor robustness. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a SLAM back-end pose optimization method based on graph neural network.

[0004] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0005] A SLAM backend posture optimization method based on graph neural network includes the following steps:

[0006] S1, using feature point detection method to extract key feature points of the image stream continuously input into the SLAM system, and record the feature descriptors of each key feature point;

[0007] S2, projecting the key feature points into three-dimensional space, obtaining corresponding points of the left camera image and the right camera image in the continuous image stream through stereo matching, obtaining matching points of the key feature points, and triangulating the matching points to obtain the three-dimensional coordinates of the key feature points;

[0008] S3, set the first frame in the image stream continuously input to the SLAM system as a key frame, and determine whether the average parallax between the current frame and the previous key frame in the image stream continuously input to the SLAM system exceeds a set threshold, if so, take the current frame as the key frame, record the key feature points extracted from the key frame and the feature descriptors of the key feature points, and establish a key frame database, otherwise, take the next frame and continue to compare it with the previous key frame;

[0009] S4, determining whether the number of key frames in the key frame database is greater than a preset threshold, if so, executing step S5, otherwise, executing step S1;

[0010] S5, defining a graph structure, and matching key feature points according to key frames, key frame feature points, and feature descriptors of key frame feature points in a key frame database, determining a common view relationship between key frames, and obtaining a graph topology structure of the key frames;

[0011] S6. Input the graph topology structure of the key frame into the graph neural network for pose optimization, and iteratively update the key frame pose using the message passing mechanism to obtain the optimized key frame pose;

[0012] S7. After the posture optimization, the key frames in the key frame database are sorted in positive order according to the size of the timestamp, and the first half of the key frames in the positive order are removed, and the second half of the key frames are used for the next round of data collection and optimization, that is, step S1 is re-executed to extract the key feature points of a new round of image streams.

[0013] Furthermore, step S1 specifically includes:

[0014] For each frame of image input into the SLAM system, the ORB feature point detection method is used to detect the corner points and edge features in each frame of image, that is, the key feature points, and the position of the key feature points is obtained, namely:

[0015] p i =(x i ,y i )

[0016] Among them, p i Indicates the position of the key feature point with index i, x i Represents the horizontal coordinate of the key feature point with index i, y i Represents the ordinate of the key feature point with index i;

[0017] The feature descriptors of key feature points are recorded at the same time.

[0018] Furthermore, step S2 specifically includes:

[0019] The key feature points extracted in step S1 are projected into the three-dimensional space, and the corresponding points of the left camera image and the right camera image in the continuous image stream, i.e., the matching points of the key feature points, are obtained through stereo matching. The positions of the matching points are triangulated using the camera's intrinsic parameter matrix to obtain the three-dimensional coordinates of the key feature points, i.e.:

[0020] P i =(X i ,Y i ,Z i )

[0021] Among them, P i represents the three-dimensional coordinates of the key feature point with index i, X i Indicates the horizontal coordinate of the key feature point with index i in three-dimensional coordinates, Y i Represents the vertical coordinate of the key feature point with index i in the three-dimensional coordinate system, Z i Represents the vertical coordinate of the key feature point with index i in three-dimensional coordinates.

[0022] Furthermore, step S5 specifically includes:

[0023] S51. Define a graph structure G=(V,E), where V represents a set of poses of key frames, and E represents a set of co-viewing relationships between key frames;

[0024] S52, according to the defined graph structure, obtain the key frames, key frame feature points and feature descriptors of the key frame feature points in the key frame database, and match the key frame feature points of the key frames to obtain the graph topology of the key frames, specifically:

[0025] If the i'th node v in the graph structure i' With the jth node v j The key frame K represents i' With K j There is a co-viewing relationship between the two key frames, that is, there are matching key feature points between the two key frames, then the i'th node v is constructed i' With the jth node v j The key frame K represents i' With K j The edge between i'j , and at the same time i'j The weight w i'j Set as keyframe K i' With K j The normalized value of the number of key feature point matches between them is used to obtain the graph topology of the key frame.

[0026] Furthermore, step S6 specifically includes:

[0027] S61, inputting the graph topology structure of the key frame obtained in step S5 into the graph neural network for pose optimization, firstly using the message passing mechanism to calculate the information transmitted from the adjacent nodes for each node in the graph topology structure of the key frame;

[0028] Among them, the information transmitted from the adjacent nodes is the influence of the neighboring nodes on the current node;

[0029] S62, according to the information transmitted from the adjacent nodes, the aggregation function is used to summarize and obtain the hidden state of each node in the next layer of the graph neural network;

[0030] S63. According to the hidden state of each node in the next layer of the graph neural network, the message transmission and aggregation of the next layer of the graph neural network are performed, and the iterative update is continued until the last layer of the graph neural network to obtain the optimized key frame pose, that is:

[0031] T * =(R * ,t * )

[0032] Among them, T * represents the optimized key frame pose, R * represents the optimized rotation matrix, t * Represents the optimized translation vector.

[0033] Furthermore, the calculation formula of the information transmitted from the adjacent nodes in step S61 is:

[0034]

[0035] Among them, m i'j represents the jth node v j Passed to the i'th node v i' The news, θ represents the message passing function, Indicates the i'th node v i' The k-th hidden state of represents the jth node v j The k-th hidden state, e i'j Indicates the i'th node v i' With the jth node v j The key frame K represents i' With K j The edge between i'j Represents edge e i'j The weight of .

[0036] Furthermore, the formula for the hidden state of each node in the next layer of the graph neural network in step S62 is:

[0037]

[0038] in, Indicates the i'th node v i' The k+1th hidden state, AGG represents the aggregation function, σ represents the activation function, Indicates the i'th node v i' The set of neighbor nodes.

[0039] The present invention has the following beneficial effects:

[0040] 1. The SLAM backend pose optimization method based on graph neural network proposed in the present invention is faster and occupies less memory when constructing large scenes;

[0041] 2. When performing end-to-end learning, key feature points and key frames are directly extracted from the original observed image stream data, and the key frame poses are optimized, which reduces the intermediate steps, simplifies the process, and makes the key frame poses more accurate, while also having stronger consistency and robustness. Therefore, the optimized key frame poses are used to update the map and path planning, which will improve the performance and accuracy of the entire SLAM system;

[0042] 3. Reduce the number of key frames by half and retain some of the key frames that have been processed. This can ensure the smooth progress of the next round of data collection and optimization, while balancing the continuity of data and the efficient use of resources, and improving the accuracy and stability of positioning and mapping. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A schematic diagram of a flow chart of a SLAM backend pose optimization method based on a graph neural network proposed in the present invention;

[0044] Figure 2 Schematic diagram of the training process of graph neural network. DETAILED DESCRIPTION

[0045] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0046] like Figure 1 As shown, a SLAM backend pose optimization method based on graph neural network includes the following steps S1-S7:

[0047] S1. The key feature points of the image stream continuously input into the SLAM system are extracted using a feature point detection method, and the feature descriptors of each key feature point are recorded.

[0048] In this embodiment, the SLAM system is a simultaneous positioning and mapping system; wherein, the present invention adopts an ORB feature point detection method to extract key feature points, but in principle, any feature point detection method can be used.

[0049] Specifically, step S1 includes:

[0050] For each frame of image input into the SLAM system, the ORB feature point detection method is used to detect the corner points and edge features in each frame of image, that is, the key feature points, and the position of the key feature points is obtained, namely:

[0051] p i =(x i ,y i )

[0052] Among them, p i Indicates the position of the key feature point with index i, x i Represents the horizontal coordinate of the key feature point with index i, y i Represents the vertical coordinate of the key feature point with index i.

[0053] The feature descriptors of key feature points are recorded at the same time.

[0054] In this embodiment, the purpose of recording the feature descriptor of each key feature point is to facilitate the description of the key feature point so as to match the key feature points between different key frames in the subsequent steps.

[0055] S2. Project the key feature points into three-dimensional space, obtain the corresponding points of the left camera image and the right camera image in the continuous image stream through stereo matching, obtain the matching points of the key feature points, and triangulate the matching points to obtain the three-dimensional coordinates of the key feature points.

[0056] Specifically, step S2 specifically includes:

[0057] The key feature points extracted in step S1 are projected into the three-dimensional space, and the corresponding points of the left camera image and the right camera image in the continuous image stream, i.e., the matching points of the key feature points, are obtained through stereo matching. The positions of the matching points are triangulated using the camera's intrinsic parameter matrix to obtain the three-dimensional coordinates of the key feature points, i.e.:

[0058] P i =(X i ,Y i ,Z i )

[0059] Among them, P irepresents the three-dimensional coordinates of the key feature point with index i, X i Indicates the horizontal coordinate of the key feature point with index i in three-dimensional coordinates, Y i Represents the vertical coordinate of the key feature point with index i in the three-dimensional coordinate system, Z i Represents the vertical coordinate of the key feature point with index i in three-dimensional coordinates.

[0060] In this embodiment, the triangulation formula is: Where b is the camera baseline, d is the parallax, and f is the focal length. represents the horizontal coordinate of the left camera image, represents the horizontal coordinate of the right camera image, represents the ordinate of the left camera image, Indicates the vertical coordinate of the right camera image.

[0061] S3, set the first frame in the image stream continuously input to the SLAM system as a key frame, and determine whether the average disparity between the current frame and the previous key frame in the image stream continuously input to the SLAM system exceeds a set threshold; if so, set the current frame as the key frame, record the key feature points extracted from the key frame and the feature descriptors of the key feature points, and establish a key frame database; otherwise, take the next frame and continue to compare it with the previous key frame.

[0062] In this embodiment, the first frame in the image stream continuously input into the SLAM system is artificially set as a key frame, so as to determine whether there is a key frame in the subsequent frame; at the same time, the key feature points extracted from the identified key frame and the feature descriptors of the key feature points are recorded to establish a key frame database; the purpose of establishing a key frame database is: when the number of key frames in the key frame database reaches a preset threshold, the optimization process is triggered in the subsequent steps to reduce the amount of calculation, that is, not every frame collected by the camera is optimized, but only the key frame is optimized. Among them, the key frame can be selected based on the change in the number of key feature points and the amount of motion between adjacent frames (such as the size of translation and rotation). In addition, the posture selected in this embodiment is the key frame (not all image frames), and its purpose is to reduce the amount of data, improve the optimization efficiency, and meet the real-time requirements.

[0063] S4. Determine whether the number of key frames in the key frame database is greater than a preset threshold. If so, execute step S5; otherwise, execute step S1.

[0064] In this embodiment, it is determined whether the number of key frames selected in step S3 is greater than a preset threshold. If so, the next step is entered for processing. Otherwise, step S1 is executed to continue extracting key feature points from the image stream and recording key frames. The purpose is to ensure that optimization is performed only when enough key frames are accumulated to ensure the stability and effect of the optimization.

[0065] S5. Define a graph structure, and perform key feature point matching based on key frames, key frame feature points, and feature descriptors of key frame feature points in a key frame database to determine the co-viewing relationship between key frames and obtain a graph topology structure of the key frames.

[0066] In this embodiment, when the number of key frames meets the preset threshold condition, a graph structure is defined, that is, G=(V, E), and the edges and weights between key frames are determined by matching key feature points, thereby obtaining a graph topology structure of the key frames.

[0067] Specifically, step S5 includes S51-S52:

[0068] S51. Define a graph structure as G=(V,E), where V represents a set of poses of key frames, and E represents a set of co-viewing relationships between key frames.

[0069] In this embodiment, the key frame pose set V is composed of the poses T of each key frame, wherein the key frame pose T includes a translation vector t and a rotation matrix R, that is, T = (R, t). The common view relationship set E between key frames is composed of the common view relationship edges e between key frames.

[0070] S52, according to the defined graph structure, obtain the key frames, key frame feature points and feature descriptors of the key frame feature points in the key frame database, and match the key frame feature points of the key frames to obtain the graph topology of the key frames, specifically:

[0071] If the i'th node v in the graph structure i' With the jth node v j The key frame K represents i' With K j There is a co-viewing relationship between the two key frames, that is, there are matching key feature points between the two key frames, then the i'th node v is constructed i' With the jth node v j The key frame K represents i' With K j The edge between i'j , and at the same time i'j The weight w i'j Set as keyframe K i' With K j The normalized value of the number of key feature point matches between them is used to obtain the graph topology of the key frame.

[0072] In this embodiment, the graph topology of the key frame obtained after the key feature point matching includes all key frames and the co-viewing relationship between the key frames, so as to provide a basis for the subsequent pose optimization. In addition, in this embodiment, the weight of the edge is set to reasonably integrate the matching information of the previous key feature points, that is, to integrate the set constraints into the subsequent graph neural network training, thereby improving the interpretability of the graph neural network and the key frame pose optimization accuracy.

[0073] S6. Input the graph topology structure of the key frame into the graph neural network for pose optimization, and use the message passing mechanism to iteratively update the key frame pose to obtain the optimized key frame pose.

[0074] The neural network model selected in this embodiment is a graph neural network (GNN), and the graph neural network is trained to obtain a trained graph neural network, so as to optimize the key frame poses in the constructed key frame graph topology structure using the trained graph neural network. The training process of the graph neural network is as follows: Figure 2 As shown in the figure, specifically: take the same number of key frames as the preset threshold as the number of nodes for graph neural network training, extract key feature points for each key frame, use these key feature point descriptors and pixel positions as the input of node features of the graph neural network, and obtain the trained graph neural network through continuous iterative updates, and output the relative poses between key frames.

[0075] In addition, before the graph topology of the key frames is input into the graph neural network for pose optimization, the connection status of the edges between the key frames has been determined. That is, whether the edge relationship is established is determined by matching the key feature points of the two frames. Therefore, when the graph topology of the key frames is input into the graph neural network for pose optimization, the edge relationship remains unchanged. The purpose is to avoid eliminating erroneous edge structures in the design of the graph neural network, making the structure of the graph neural network simpler and less complex.

[0076] Specifically, step S6 specifically includes S61-S63:

[0077] S61. Input the graph topology structure of the key frame obtained in step S5 into the graph neural network for posture optimization. First, use the message passing mechanism to calculate the information transmitted from the adjacent nodes for each node in the graph topology structure of the key frame; wherein, the information transmitted from the adjacent nodes is the influence of the neighboring nodes on the current node.

[0078] Specifically, the calculation formula of the information transmitted from the adjacent nodes in step S61 is:

[0079]

[0080] Among them, m i'jrepresents the jth node v j Passed to the i'th node v i' The news, θ represents the message passing function, Indicates the i'th node v i' The k-th hidden state of represents the jth node v j The k-th hidden state, e i'j Indicates the i'th node v i' With the jth node v j The key frame K represents i' With K j The edge between i'j Represents edge e i'j The weight of .

[0081] S62. Based on the information transmitted from adjacent nodes, the aggregation function is used to summarize and obtain the hidden state of each node in the next layer of the graph neural network.

[0082] Specifically, the formula for the hidden state of each node in the next layer of the graph neural network in step S62 is:

[0083]

[0084] in, Indicates the i'th node v i' The k+1th hidden state, AGG represents the aggregation function, σ represents the activation function, Indicates the i'th node v i' The set of neighbor nodes.

[0085] S63. According to the hidden state of each node in the next layer of the graph neural network, the message transmission and aggregation of the next layer of the graph neural network are performed, and the iterative update is continued until the last layer of the graph neural network to obtain the optimized key frame pose, that is:

[0086] T * =(R * ,t * )

[0087] Among them, T * represents the optimized key frame pose, R * represents the optimized rotation matrix, t * Represents the optimized translation vector.

[0088] In this embodiment, after processing by the graph neural network, the optimized positions of all key frames in the time period are output to obtain optimized key frame poses. These optimized key frame poses are not only more accurate, but also have stronger consistency and robustness. Therefore, using the optimized key frame poses for updating maps and path planning will improve the performance and accuracy of the entire SLAM system.

[0089] S7. After the posture optimization, the key frames in the key frame database are sorted in positive order according to the size of the timestamp, and the first half of the key frames in the positive order are removed, and the second half of the key frames are used for the next round of data collection and optimization, that is, step S1 is re-executed to extract the key feature points of a new round of image streams.

[0090] In the present embodiment, after outputting the optimization result, the currently constructed graph topology is destroyed to release memory and resources. In order to balance the continuity of data and the efficient use of resources, the number of key frames is halved, and a part of the key frames that have been processed is retained, that is, the second half of the key frames in the key frames that are sorted in positive order are retained, so that the smooth progress of the next round of data collection and optimization can be ensured. Then return to step S1, restart to extract feature points and record key frames from the image stream, and gradually build a new graph structure, through such a cycle, the SLAM system can continuously optimize the posture, improve the accuracy and stability of positioning and mapping. In addition, steps S1-S7 fully illustrate that the present invention is segmented optimization, that is, according to the situation that new image streams are constantly flowing in the real-time scene, it is designed to accumulate to a certain image frame, and then the image frame of the period is optimized, and its purpose is more flexible, and the situation that the image stream is always input can be processed, and the posture optimization is performed in a small range while collecting the image stream.

[0091] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0092] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A SLAM backend pose optimization method based on graph neural network, characterized in that: The following steps are involved: S1, using feature point detection method to extract key feature points of the image stream continuously input into the SLAM system, and record the feature descriptors of each key feature point; S2, projecting the key feature points into three-dimensional space, obtaining corresponding points of the left camera image and the right camera image in the continuous image stream through stereo matching, obtaining matching points of the key feature points, and triangulating the matching points to obtain the three-dimensional coordinates of the key feature points; S3, set the first frame in the image stream continuously input to the SLAM system as a key frame, and determine whether the average parallax between the current frame and the previous key frame in the image stream continuously input to the SLAM system exceeds a set threshold, if so, take the current frame as the key frame, record the key feature points extracted from the key frame and the feature descriptors of the key feature points, and establish a key frame database, otherwise, take the next frame and continue to compare it with the previous key frame; S4, determining whether the number of key frames in the key frame database is greater than a preset threshold, if so, executing step S5, otherwise, executing step S1; S5, defining a graph structure, and matching key feature points according to key frames, key frame feature points, and feature descriptors of key frame feature points in a key frame database, determining a common view relationship between key frames, and obtaining a graph topology structure of the key frames; S6. Input the graph topology structure of the key frame into the graph neural network for pose optimization, and iteratively update the key frame pose using the message passing mechanism to obtain the optimized key frame pose; S7. After the posture optimization, the key frames in the key frame database are sorted in positive order according to the size of the timestamp, and the first half of the key frames in the positive order are removed, and the second half of the key frames are used for the next round of data collection and optimization, that is, step S1 is re-executed to extract the key feature points of a new round of image streams.

2. The SLAM backend posture optimization method based on graph neural network according to claim 1, characterized in that: Step S1 specifically includes: For each frame of image input into the SLAM system, the ORB feature point detection method is used to detect the corner points and edge features in each frame of image, that is, the key feature points, and the position of the key feature points is obtained, namely: p i =(x i ,y i ) Among them, p i Indicates the position of the key feature point with index i, x i Represents the horizontal coordinate of the key feature point with index i, y i Represents the ordinate of the key feature point with index i; The feature descriptors of key feature points are recorded at the same time.

3. The SLAM back-end posture optimization method based on graph neural network according to claim 1, characterized in that, Step S2 specifically includes: The key feature points extracted in step S1 are projected into the three-dimensional space, and the corresponding points of the left camera image and the right camera image in the continuous image stream, i.e., the matching points of the key feature points, are obtained through stereo matching. The positions of the matching points are triangulated using the camera's intrinsic parameter matrix to obtain the three-dimensional coordinates of the key feature points, i.e.: P i =(X i ,Y i ,Z i ) Among them, P i represents the three-dimensional coordinates of the key feature point with index i, X i Indicates the horizontal coordinate of the key feature point with index i in three-dimensional coordinates, Y i Represents the vertical coordinate of the key feature point with index i in the three-dimensional coordinate system, Z i Represents the vertical coordinate of the key feature point with index i in three-dimensional coordinates.

4. The SLAM backend posture optimization method based on graph neural network according to claim 1, characterized in that: Step S5 specifically includes: S51. Define a graph structure G=(V,E), where V represents a set of poses of key frames, and E represents a set of co-viewing relationships between key frames; S52, according to the defined graph structure, obtain the key frames, key frame feature points and feature descriptors of the key frame feature points in the key frame database, and match the key frame feature points of the key frames to obtain the graph topology of the key frames, specifically: If the i'th node v in the graph structure i' With the jth node v j The key frame K represents i' With K j There is a co-viewing relationship between the two key frames, that is, there are matching key feature points between the two key frames, then the i'th node v is constructed i' With the jth node v j The key frame K represents i' With K j The edge between i'j , and at the same time i'j The weight w i'j Set as keyframe K i' With K j The normalized value of the number of key feature point matches between them is used to obtain the graph topology of the key frame.

5. The SLAM backend posture optimization method based on graph neural network according to claim 1, characterized in that: Step S6 specifically includes: S61, inputting the graph topology structure of the key frame obtained in step S5 into the graph neural network for pose optimization, firstly using the message passing mechanism to calculate the information transmitted from the adjacent nodes for each node in the graph topology structure of the key frame; Among them, the information transmitted from the adjacent nodes is the influence of the neighboring nodes on the current node; S62, according to the information transmitted from the adjacent nodes, the aggregation function is used to summarize and obtain the hidden state of each node in the next layer of the graph neural network; S63. According to the hidden state of each node in the next layer of the graph neural network, the message transmission and aggregation of the next layer of the graph neural network are performed, and the iterative update is continued until the last layer of the graph neural network to obtain the optimized key frame pose, that is: T * =(R * ,t * ) Among them, T * represents the optimized key frame pose, R * represents the optimized rotation matrix, t * Represents the optimized translation vector.

6. The SLAM backend posture optimization method based on graph neural network according to claim 5, characterized in that: The calculation formula of the information transmitted from the adjacent nodes in step S61 is: Among them, m i'j represents the jth node v j Passed to the i'th node v i' The news, θ represents the message passing function, Indicates the i'th node v i' The k-th hidden state of represents the jth node v j The k-th hidden state, e i'j Indicates the i'th node v i' With the jth node v j The key frame K represents i' With K j The edge between i'j Represents edge e i'j The weight of .

7. The SLAM backend posture optimization method based on graph neural network according to claim 6, characterized in that: The formula for the hidden state of each node in the next layer of the graph neural network in step S62 is: in, Indicates the i'th node v i' The k+1th hidden state, AGG represents the aggregation function, σ represents the activation function, Indicates the i'th node v i' The set of neighbor nodes.

Images