A loop closure detection method based on point cloud semantic graph descriptor and position information

Abstract

The application relates to a loop detection method based on a point cloud semantic graph descriptor and position information, comprising the following steps: acquiring laser point cloud data and satellite positioning data of a current frame, performing semantic segmentation processing on the laser point cloud data, converting into a preset point cloud semantic graph descriptor form, screening a loop candidate frame from historical frame data through distance matching and node matching; converting the current frame and the loop candidate frame into a circular ring graph form, and attempting to obtain a target loop frame through similarity verification and geometric consistency verification. Compared with the prior art, the application solves the problems of difficult loop detection, large loop matching calculation amount and low loop matching accuracy in a mobile robot SLAM system, and improves the positioning accuracy and system robustness of the mobile robot.

Description

Technical Field

[0001] This invention relates to the field of map construction technology, and in particular to a loop closure detection method based on point cloud semantic graph descriptors and location information. Background Technology

[0002] With the continuous advancement of computer technology and sensor technology, mobile robots are also developing rapidly. Positioning is an important part of the mobile robot technology system, and its accuracy and real-time performance directly determine the operation of the entire mobile robot.

[0003] Simultaneous Localization and Mapping (SLAM) is a key technology for mobile robot localization. LiDAR SLAM utilizes a 3D lidar system and the robot's own state to achieve mobile robot localization. Loop closure detection aims to find the robot's past traversal points. Adding a loop closure detection module to the SLAM system reduces the cumulative error of the SLAM system; it is a crucial module in SLAM.

[0004] Currently, most SLAM methods use loop closure detection to detect loops by determining whether the pose distance between the current frame and historical frames is within a threshold range. This type of method is very time-consuming, requiring matching of all point cloud data that meet the distance conditions and identifying loops. The low efficiency of loop closure detection leads to a decrease in the positioning accuracy of the mobile robot, and may even cause the experiment to fail.

[0005] Chinese patent application CN202211042843.9 provides a loop closure detection method based on point cloud semantic laser SLAM. The method comprises the following steps: S1, any two frames of point cloud scan data, denoted as the first frame and the second frame; S2, inputting the first frame and the second frame into a semantic segmentation network for point cloud semantic segmentation, and adding corresponding labels to the semantic blocks; S3, dividing the first frame and the second frame into M parts according to the horizontal scanning distance of the points, and describing the semantic information using the semantic vectors of the first frame and the second frame respectively; S4, obtaining frames where the 1-norm difference between the semantic vectors of the first frame and the second frame is less than a preset value, as loop closure candidate frames; S5, inputting the loop closure candidate frames as parameters into a registration algorithm to construct loop closure constraints.

[0006] The aforementioned application obtains semantic information from point cloud data and divides it equally, then uses norms to filter candidate frames for loop closure, thereby reducing the computational load between two image frames. However, using norms for filtering in the aforementioned application results in the loss of relative positional information in 3D space, leading to unsatisfactory loop closure matching accuracy. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a loop closure detection method based on point cloud semantic graph descriptors and location information, so as to improve the success rate of loop closure matching while reducing the amount of computation.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] This invention provides a loop closure detection method based on point cloud semantic graph descriptors and location information, comprising the following steps:

[0010] The laser point cloud data and satellite positioning data of the current frame are acquired. The laser point cloud data is semantically segmented and converted into a preset point cloud semantic graph descriptor. Loop closure candidate frames are selected from historical frame data through distance matching and node matching.

[0011] The current frame and the candidate loop closure frames are converted into a circular graph, and the target loop closure frame is obtained through similarity verification and geometric consistency verification.

[0012] As a preferred technical solution, the process of semantic segmentation processing of the laser point cloud data includes the following steps:

[0013] Semantic segmentation is performed on the preprocessed laser point cloud data to extract objects in the environment. An undirected graph is used to construct semantic description information of the point cloud. In the undirected graph, nodes represent the positions of objects and edges represent the distances between objects.

[0014] As a preferred technical solution, the objects in the environment include vehicles, tree trunks, and poles.

[0015] As a preferred technical solution, the distance matching specifically includes the following steps:

[0016] Based on the satellite positioning data of the current frame, historical frames in which the distance between the positioning location and the positioning location of the current frame is greater than or less than a preset threshold are selected.

[0017] As a preferred technical solution, the node matching specifically includes the following steps:

[0018] For the historical frames obtained through distance matching, based on the laser point cloud data in the form of the current frame point cloud semantic graph descriptor, historical frames with a number of environmental objects close to the number of environmental objects in the current frame are selected as the loop closure candidate frames.

[0019] As a preferred technical solution, for any frame among the current frame and the loop closure candidate frames, the process of generating the circular graph includes the following steps:

[0020] Construct multiple concentric rings, the number of which is the same as the number of edge types in the undirected graph obtained from semantic segmentation;

[0021] Starting from the same straight line, each concentric ring is divided into multiple uniform sectors. The Euclidean distances of the same type of edges in the environment are calculated and sorted. The number of sectors is allocated according to the magnitude of the Euclidean distance to realize the construction of the ring graph.

[0022] As a preferred technical solution, the similarity verification process includes the following steps:

[0023] Based on the current frame and loop closure candidate frames in the form of a ring graph, calculate the ring similarity between the current frame and each loop closure candidate frame, and determine whether the ring similarity of the loop closure candidate frames is greater than or less than a preset threshold. If it is, retain it; otherwise, discard it.

[0024] As a preferred technical solution, the geometric consistency verification process includes the following steps:

[0025] The following formula is used to calculate the loop closure candidate frames after similarity matching and the loop closure detection loss value between the current frame. It is then determined whether the loop closure detection loss value of the candidate frame is greater than or less than a preset threshold. If so, it is selected as the target loop closure frame; otherwise, it is discarded.

[0026]

[0027] Where loss is the loop closure detection loss value, p l For the current frame's point cloud set, p h Let n be the set of point clouds for the candidate frames of the loop closure, where n represents the number of 3D points in the current data frame point cloud, T represents the transformation matrix, and p il and p ih These represent the i-th 3D point in the current frame and the loopback candidate frame, respectively.

[0028] As a preferred technical solution, the loop closure detection method further includes the following steps:

[0029] Determine whether the target loopback frame has been successfully matched. If so, the transformation matrix T is the initial pose of the six degrees of freedom in the SLAM system. If not, store the satellite positioning data and laser point cloud data in the form of point cloud semantic graph descriptors of the current frame into a preset point cloud semantic graph container.

[0030] As a preferred technical solution, the laser point cloud data is acquired by a three-dimensional lidar, and the satellite positioning data is acquired by a GNSS antenna and a GNSS receiver.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] (1) High success rate of loop closure matching and low computational cost: Unlike existing schemes that use norms to filter candidate loop closure frames, this invention first acquires the laser point cloud data and satellite positioning data of the current frame, and performs semantic segmentation on the laser point cloud data. Through distance matching and node matching, loop closure candidate frames are selected from historical frame data to complete the first stage of matching. After that, the current frame and loop closure candidate frames are converted into a donut graph and the target loop closure frame is obtained through similarity matching and geometric consistency matching to achieve the second stage of matching. This method integrates Euclidean distance, number of nodes, donut graph similarity and loop closure detection loss to effectively filter historical frames, making full use of the high-level constraints of features. By identifying the relative position of objects in three-dimensional space, the entire scene can be understood from a macroscopic perspective. This invention can perform fast, efficient and accurate loop closure detection, which can effectively improve the positioning accuracy and robustness of the SLAM system. Common and stable objects in outdoor scenes are used as target objects for semantic segmentation to construct cloud semantic graph descriptors, which improves the representativeness of point cloud descriptors.

[0033] (2) High compatibility and low cost: This invention can realize the loop closure detection function using only three-dimensional lidar and GNSS system. It can be used in the current mainstream SLAM system. The equipment cost is low and the operation is simple and easy to learn. Attached Figure Description

[0034] Figure 1 This is a flowchart of the loop closure detection method based on point cloud semantic graph descriptors and location information in the embodiment;

[0035] Figure 2 This is a schematic diagram of a point cloud semantic graph descriptor.

[0036] Figure 3 A schematic diagram illustrating the method for constructing semantic graph descriptors for point clouds;

[0037] Figure 4 This is a schematic diagram of a donut diagram;

[0038] Figure 5 This is a schematic diagram of the method for constructing a donut graph. Detailed Implementation

[0039] 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, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0040] Example 1

[0041] To address the aforementioned problems in the prior art, the present invention aims to overcome the shortcomings of existing technologies and provide a loop closure detection method based on point cloud semantic graph descriptors. Furthermore, it incorporates location information obtained from the Global Navigation Satellite System (GNSS) during the loop closure detection process. This addresses the challenges of difficult loop closure detection, high computational load for loop closure matching, and low accuracy in mobile robot SLAM systems, thereby improving the positioning accuracy and system robustness of mobile robots.

[0042] like Figure 1 This embodiment provides a loop closure detection method based on point cloud semantic graph descriptors and location information, including the following steps:

[0043] Step S1: Install equipment and acquire 3D laser point cloud and GNSS data:

[0044] The Velodyne VLP-16 3D LiDAR, GNSS antenna, and GNSS receiver are all mounted on the roof of the vehicle. This 3D laser scanning range is 100 meters, with 16 3D LiDARs within a 30° field of view. The vertical resolution is 2°, and the horizontal resolution is 0.2° at a scanning frequency of 10Hz. Real-Time Kinematic (RTK) technology can be used to obtain centimeter-level positional information. The 3D LiDAR and GNSS receiver transmit the collected point cloud data and GNSS-RTK data to the onboard industrial control computer. The onboard industrial control computer and the autonomous vehicle exchange data and send commands through a Robot Operating System to obtain the input of 3D laser point cloud data and GNSS-RTK data.

[0045] Step S2, construct the point cloud semantic graph descriptor container:

[0046] First, the acquired 3D point cloud data is preprocessed to remove motion distortion and downsample the data to reduce its size. The generation method for the point semantic graph descriptor obtained from the radar point cloud data frame conversion involves the following steps:

[0047] RangeNet++ was used to perform semantic segmentation on point cloud data, extracting objects from the environment to construct point cloud semantic graph descriptors. To avoid the point cloud descriptors being unrepresentative due to high randomness of target objects, common and stable objects in outdoor scenes were selected as the target objects for semantic segmentation. For example, the selected target objects are vehicles, tree trunks, and poles.

[0048] For each candidate frame of point cloud data, after semantic segmentation of the point cloud to identify the target object using RangeNet++, an undirected graph G = <V, E> is used to construct the point cloud descriptor. Node V represents the position of the target object in the undirected graph G (using the centroid of the target object to represent its position in space), E represents the edges between nodes V, and edge e represents the position of the target object in space. ij = <V i V j > for V i V j The Euclidean distance between two nodes.

[0049] In an undirected graph G, the target objects can be paired to obtain 3 types of vertices and 6 types of edges. V has three types of vertices: vehicles, tree trunks, and poles; E has six types of edges: pole-pole, tree trunk-tree trunk, vehicle-vehicle, vehicle-tree trunk, tree trunk-pole, and pole-vehicle.

[0050] like Figure 2 The above is a schematic diagram of a point cloud semantic graph descriptor, such as... Figure 3 The above is a schematic diagram of a point cloud semantic graph descriptor construction method. The descriptor and location information corresponding to the key frame are stored together in the point cloud semantic graph descriptor container in the form of key-value pairs. The container uses the Vector container in C++ for storage. Subsequently, when performing loop closure matching, all the descriptors in the container are used to construct a kd-tree and use the standard k-nearest neighbor search algorithm to quickly search for loop closure candidate frames.

[0051] Step S3, Distance Matching:

[0052] When performing loop closure detection, firstly, based on the GNSS location information corresponding to the current frame, search for the historical key frame that is closest to the location of the current frame's point cloud descriptor in the point cloud semantic graph descriptor container obtained in step S3. Record the historical key frames that meet the distance condition as loop closure candidate frames. The distance threshold is set to 0.25m. This invention sets up 10 loop closure candidate frames.

[0053] Step S4, Node Matching:

[0054] After obtaining the loop closure candidate frames in step S3, a preliminary matching is performed on the number of three types of target object nodes in the point cloud semantic graph descriptor corresponding to the loop closure candidate frames. Loop closure candidate frames whose total number of the three types of target objects is within a predetermined threshold range are included in the candidate objects. Let the current number of object nodes be V, and the number of object nodes in the loop closure candidate frames to be matched be V. total For loop-loop candidate frames whose sum of the three target objects falls within a predetermined threshold range λ, we include them in the candidate objects, where λ is set to 15. total The following formula should be satisfied:

[0055] Vtotal =[V-15,V+15]

[0056] Step S5, the first layer of verification in the loop closure—data frame similarity matching:

[0057] After obtaining the candidate objects in step S4, they are converted into point cloud semantic graph descriptors and a donut graph is generated simultaneously. The similarity score of the two data frames is calculated using the donut graphs of the loop closure candidate frame and the current frame, so as to determine whether the loop closure candidate frame and the current frame constitute a loop closure frame.

[0058] like Figure 5 The specific method for generating a donut chart consists of the following steps:

[0059] In steps S51-S52, the number of rings in the anodic graph is the same as the number of types of E in the undirected graph G. The anodic graph in this paper is 6-ringed, corresponding to the six edge types in E.

[0060] This invention only introduces one of the six edge types in E—E. trunk-pole The process of constructing a ring in the corresponding ring graph is the same for the other five edge types.

[0061] Step S53, assume there are m E's in the point cloud semantic graph descriptor. trunk-pole Side, these m E trunk-pole The edges form an edge set S. trunk-pole S trunk-pole The range of Euclidean distances to the edges in the interval is:

[0062]

[0063] In the above formula, s min It is S trunk-pole Minimum Euclidean distance between the middle and sides, s max It is S trunk-pole The maximum value of the Euclidean distance between the middle and sides. Step S54, for S trunk-pole Sort each edge in ascending order according to its Euclidean distance to obtain an ordered set of edges S. trunk-pole′ for:

[0064]

[0065] Step S55, S trunk-pole' The corresponding annulus is divided into ms max There are 1 sector, and the angle corresponding to each sector is 1. Then the ordered edge set S trunk-pole' The number of sectors corresponding to each edge in the equation N i for:

[0066]

[0067] S trunk-pole' The number of sectors corresponding to each side is arranged sequentially in the ring. One sector is used to separate the sectors corresponding to each side. Starting from the horizontal right line, the sectors are arranged counterclockwise for each side.

[0068] This will transfer the S in the point cloud descriptor trunk-pole' The edge of type is represented by a ring in the annulus graph. We will use S trunk-pole' The corresponding ring is denoted as C. trunk-pole Then the annulus C trunk-pole Its sector set in the donut diagram is:

[0069] C trunk-pole ={N0,N1,...,N m-1}

[0070] Then, by representing the remaining five types of edges in the point cloud descriptor with rings in the dodecahedron, a complete dodecahedron C is formed. graph The generated donut chart is as follows Figure 4 As stated above.

[0071] C graph =(C trunk-trunk ,...C pole-vehicle )

[0072] Step S56, the dome graph corresponding to such a point cloud semantic graph descriptor is C. graph We can use the loop closure candidate frame and the donut graph of the current frame to perform similarity matching, which can be used to quickly filter loop closure candidate frames and thus obtain loop closure frames.

[0073] When performing ring graph matching, the ring graphs of the candidate ring frames and the current frame are first aligned by their centers and horizontal right lines. Then, the ring similarity matching is calculated by matching the similarity of the edges of each ring in the ring graph at different Euclidean distances. The ring similarity matching formula is as follows:

[0074]

[0075] In the above formula, This represents the donut chart corresponding to the current frame. This represents the annulus graph corresponding to a certain loop closure candidate frame, where m is the number of annulus in the annulus graph (i.e., the types of edges in the point cloud semantic graph descriptor).

[0076] The threshold α is used to determine whether a loop closure might exist between the current frame and the loop closure candidate frame.

[0077]

[0078] If the similarity score is less than the set threshold α (in this embodiment, the threshold α is 0.85), it means that no loopback frame has been matched. The descriptor generated for the current frame and its corresponding GNSS-RTK location are stored together in the point cloud semantic graph descriptor container. Otherwise, it is considered that there may be a loopback between the current frame and the loopback candidate frame. After completing the first verification of the loopback, the second verification of the loopback—geometric verification—is required.

[0079] Step S6, the second verification of the loop closure—geometric verification:

[0080] After determining that there may be loops between the current frame and several loop closure candidate frames obtained in step S5, further geometric verification is required. The focus of this step is to find a geometrically consistent correspondence for the current frame, thereby determining the loop closure frame and obtaining the initial pose of the six degrees of freedom.

[0081] This invention uses the following formula to evaluate the loss for loop closure detection in candidate frames and the current frame:

[0082]

[0083] In the above formula, p l For the current frame's point cloud set, p h Let n be the set of point clouds for the candidate frames of the loop closure, where n represents the number of point clouds in the current data frame, T represents the transformation matrix, and p il and p ih These represent the i-th 3D point in the current frame and the loopback candidate frame, respectively.

[0084]

[0085] If the loss value of a loop closure candidate frame is greater than the threshold β (in this embodiment, the threshold β is 0.3), it means that no loop closure frame has been matched. The descriptor generated for the current frame and its corresponding GNSS-RTK position are stored together in the point cloud semantic graph descriptor container. Otherwise, it is determined that there is a loop in these two places. When the loop is confirmed to exist, the transformation matrix T is used as the initial pose of the six degrees of freedom.

[0086] After obtaining the loop closure frame, a loop closure constraint factor can be constructed with the current frame and added to the global factor graph model of the SLAM system. At the same time, the six-degree-of-freedom initial pose obtained during the loop closure detection process is beneficial to the point cloud matching of the front end, which can effectively reduce the cumulative error of the SLAM system and achieve real-time high-precision positioning of the mobile robot.

[0087] Compared with the prior art, the present invention has the following obvious and prominent substantive features and significant advantages:

[0088] 1. This invention solves the problems of difficult loop closure detection, large computational load for loop closure matching, and low accuracy of loop closure matching. Compared with the traditional method of using whether the pose distance between the current frame and historical frames is within a threshold range, this invention makes full use of advanced feature constraints. By identifying the relative position of objects in three-dimensional space, it understands the entire scene from a macroscopic perspective. This invention can perform loop closure detection quickly, efficiently, and accurately, and can effectively improve the positioning accuracy and robustness of the SLAM system.

[0089] 2. This invention can achieve loop closure detection using only a three-dimensional lidar and GNSS system, and can be used in current mainstream SLAM systems. The equipment is low-cost and easy to operate.

[0090] 3. The method of the present invention is simple, easy to implement, and low in cost, making it suitable for widespread use.

[0091] Example 2

[0092] Compared with Example 1, in step S2 of this example, when removing point cloud motion distortion, the motion state of the lidar can be recovered using the observation information of the Inertial Measurement Unit (IMU) and the high-frequency IMU observation propagation model, which can better correct the motion distortion of the point cloud. The specific steps are as follows:

[0093] Step 1, Install equipment and acquire 3D laser point cloud and GNSS data:

[0094] This embodiment uses an autonomous vehicle as the application object of the mobile robot. The three-dimensional LiDAR and GNSS antenna are installed on the top of the vehicle, and the GNSS receiver is installed in the trunk of the vehicle. The three-dimensional LiDAR and GNSS will collect point cloud data of the surrounding environment and obtain location information during the vehicle's driving process. The obtained three-dimensional point cloud data and location information are transmitted to the computer and saved for subsequent algorithm processing.

[0095] Step 2, construct the point cloud semantic graph descriptor container:

[0096] First, the acquired 3D point cloud data is preprocessed. A high-frequency IMU observation propagation model is used to recover the motion state of the lidar, remove motion distortion of the point cloud, and downsample the point cloud data to reduce its size. For the point semantic graph descriptor obtained by converting the radar point cloud data frame, the descriptor and position information corresponding to the key frame are stored together in the point cloud semantic graph descriptor container in the form of key-value pairs.

[0097] Step 3, Distance Matching:

[0098] When performing loop closure detection, firstly, based on the GNSS location information corresponding to the current frame, search the point cloud semantic graph descriptor container obtained in step 3 for the historical key frame that is closest to the current frame's point cloud descriptor location, and record the historical key frame that meets the distance condition as a loop closure candidate frame.

[0099] Step 4, Node Matching:

[0100] After obtaining the loop closure candidate frames in step 3, the number of three types of target object nodes in the point cloud semantic graph descriptor corresponding to the loop closure candidate frames is initially matched, and loop closure candidate frames whose total number of three types of target objects is within a predetermined threshold range are included in the candidate objects.

[0101] Step 5, the first layer of verification for loop closure—data frame similarity matching:

[0102] After obtaining the candidate objects in step 4, they are converted into point cloud semantic graph descriptors and a donut graph is generated at the same time. The similarity scores of the two data frames are calculated using the donut graphs of the loop closure candidate frame and the current frame, so as to determine whether the loop closure candidate frame and the current frame constitute a loop closure frame.

[0103] Step 6, the second verification of the loop—geometric verification:

[0104] After determining that there may be loops between the current frame and several loop closure candidate frames obtained in step 5, further geometric verification is required. The focus of this step is to find a geometrically consistent correspondence for the current frame, thereby determining the loop closure frame and obtaining the initial pose of the six degrees of freedom.

[0105] Example 3

[0106] This embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing the loop closure detection method based on point cloud semantic graph descriptors and location information as described in either Embodiment 1 or Embodiment 2.

[0107] Example 4

[0108] This embodiment provides a computer-readable storage medium including one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for performing a loop closure detection method based on point cloud semantic graph descriptors and location information as described in either Embodiment 1 or Embodiment 2.

[0109] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A loop closure detection method based on point cloud semantic graph descriptors and location information, characterized in that, Includes the following steps: The laser point cloud data and satellite positioning data of the current frame are acquired. The laser point cloud data is semantically segmented and converted into a preset point cloud semantic graph descriptor. Loop closure candidate frames are selected from historical frame data through distance matching and node matching. The current frame and the candidate loop closure frames are converted into a pie chart, and the target loop closure frame is obtained through similarity verification and geometric consistency verification. The process of semantic segmentation processing of the laser point cloud data includes the following steps: Semantic segmentation is performed on the preprocessed laser point cloud data to extract objects in the environment. An undirected graph is then used to construct a semantic description of the point cloud, where nodes represent the positions of objects and edges represent the distances between objects. Objects in the environment include vehicles, tree trunks, and poles. For any frame among the current frame and the loop closure candidate frames, the process of generating the circular graph includes the following steps: Construct multiple concentric rings, the number of which is the same as the number of edge types in the undirected graph obtained from semantic segmentation; Starting from the same straight line, each concentric ring is divided into multiple uniform sectors. The Euclidean distances of the same type of edges in the environment are calculated and sorted. The number of sectors is then allocated according to the magnitude of the Euclidean distances to construct the ring graph. The similarity verification process includes the following steps: Based on the current frame and loop closure candidate frames in the form of a ring graph, calculate the ring similarity between the current frame and each loop closure candidate frame. Determine whether the ring similarity of the loop closure candidate frames is greater than or less than a preset threshold. If it is, retain the frame; otherwise, discard it. The geometric consistency verification process includes the following steps: The following formula is used to calculate the loop closure candidate frames after similarity matching and the loop closure detection loss value between the current frame. It is then determined whether the loop closure detection loss value of the candidate frame is greater than or less than a preset threshold. If so, it is selected as the target loop closure frame; otherwise, it is discarded. in, loss The loop closure detection loss value is... p l For the current frame's point cloud set, p h This is the set of point clouds for the candidate frames of the loop closure. n This indicates the number of 3D points in the current data frame point cloud. T Represents the transformation matrix. p il and p ih These represent the current frame and the loopback candidate frame, respectively. i A 3D point, The loop closure detection method further includes the following steps: Determine if the target loopback frame was successfully matched; if so, then transform the matrix. T This is the initial pose of the six degrees of freedom in the SLAM system. Otherwise, the satellite positioning data of the current frame and the laser point cloud data in the form of point cloud semantic graph descriptors are stored in a preset point cloud semantic graph container.

2. The loop closure detection method based on point cloud semantic graph descriptors and location information according to claim 1, characterized in that, The distance matching specifically includes the following steps: Based on the satellite positioning data of the current frame, historical frames in which the distance between the positioning location and the positioning location of the current frame is greater than or less than a preset threshold are selected.

3. The loop closure detection method based on point cloud semantic graph descriptors and location information according to claim 1, characterized in that, The node matching specifically includes the following steps: For the historical frames obtained through distance matching, based on the laser point cloud data in the form of the current frame point cloud semantic graph descriptor, historical frames with a number of environmental objects close to the number of environmental objects in the current frame are selected as the loop closure candidate frames.

4. The loop closure detection method based on point cloud semantic graph descriptors and location information according to claim 1, characterized in that, The laser point cloud data is acquired using a three-dimensional lidar, and the satellite positioning data is acquired using a GNSS antenna and a GNSS receiver.

Images