SLAM map topology evaluation method and device, electronic equipment and storage medium

Abstract

The application relates to the technical field of mapping, in particular to a SLAM map topology evaluation method and device, electronic equipment and a storage medium, wherein the method comprises the following steps: acquiring point cloud data of a target area; inputting the point cloud data into a trained deep learning point cloud semantic segmentation model, performing plane detection on the point cloud of a pre-defined static scene semantic category to obtain a candidate plane set and corresponding estimated normal information; constructing an adjacency matrix of candidate planes in the candidate plane set through a parallel plane detection condition, and calculating coplanar point variance; performing normal condensation clustering on the normal information to obtain a plurality of normal sub-clusters, calculating a normal entropy mean value based on the normal of each sub-cluster, and determining a map topology evaluation result of a simultaneous localization and mapping (SLAM) map of the target area according to the coplanar point variance and / or the normal entropy mean value. Thus, the problems that SLAM pose estimation is limited by the environment and a measuring device, and the precision estimation method is limited in the related art are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mapping, in particular to a SLAM map topology evaluation method and device, an electronic device and a storage medium. BACKGROUND

[0002] With the continuous progress of sensors, mobile chips and artificial intelligence technology, intelligent robots have gradually been popularized in daily life and work scenes, and have the potential to replace humans from repetitive, low-end, simple, and heavy work. In the scenes of unmanned cleaning, transportation, security, monitoring, planting and picking, the application of intelligent unmanned work of mobile robots has unprecedented economic value and social application value.

[0003] Among them, the understanding of complex dynamic scenes and self-positioning of mobile robots is the basis for ensuring the safety of robots and subsequent tasks. The current mainstream robot autonomous positioning technology mainly adopts Simultaneous Localization And Mapping (SLAM), and the SLAM algorithm based on open-loop control can reduce the dependence of robots on additional equipment and environmental infrastructure as much as possible, and the mapping results can be used for environmental measurement, visual interaction and positioning of other robots, and are suitable for robot autonomous positioning and mapping in large-scale, Global Positioning System (GPS) limited scenes.

[0004] The existing SLAM development mainly depends on relative pose estimation and absolute pose estimation based on reference estimation metrics, that is, the relative or absolute error of the robot motion pose estimated by the specific SLAM algorithm and the actual robot motion pose change is calculated, and the error is reduced as much as possible. The reference value of the actual robot pose is obtained or the accurate differential GPS positioning information is adopted, or the high-precision measuring equipment such as laser tracker or total station is used. Among them, the differential GPS positioning information has errors and cannot support indoor environment positioning due to the influence of canyon effect and building shielding, while the laser tracker is based on active laser tracking of the assembled target ball, and can only support tracking within the line of sight within a limited distance. In large and complex shielding scenes, a large number of devices are needed for collaborative tracking, which is extremely expensive and difficult to deploy on a large scale. Other radio frequency, Ultra Wide Band (UWB) or millimeter wave positioning technologies are mainly based on deploying anchor points by modifying the environment, and estimating the target coordinates by means of triangulation positioning. The accuracy fluctuates greatly, and the attitude cannot be estimated. SUMMARY

[0005] The application provides a SLAM map topology evaluation method and device, electronic equipment and storage medium, to solve the problem that the SLAM pose estimation is limited by the environment and the measuring device, and the precision estimation needs a benchmark value.

[0006] The first aspect of the application provides a SLAM map topology evaluation method, comprising the following steps: obtaining point cloud data of a target area; inputting the point cloud data into a trained deep learning point cloud semantic segmentation model to output semantic label information of the point cloud, determining a semantic category of the point cloud based on the semantic label information, performing plane detection on the point cloud of a predefined static scene semantic category to obtain a candidate plane set and corresponding estimated normal information; constructing an adjacency matrix of candidate planes in the candidate plane set through parallel plane detection conditions, detecting subgroups belonging to coplanar planes using the adjacency matrix and a group search algorithm, calculating coplanar point variance based on the subgroups; performing normal agglomerative clustering on the normal information to obtain a plurality of normal subclusters, calculating normal entropy for the normal of each subcluster, calculating the mean of the normal entropy of all subclusters, and determining a map topology evaluation result of a SLAM map of the target area according to the coplanar point variance and / or the mean of the normal entropy.

[0007] Optionally, before inputting the point cloud data into the trained deep learning point cloud semantic segmentation model, the method further comprises: collecting point cloud data of an environment; constructing a point cloud semantic segmentation dataset based on the point cloud data with artificial point cloud semantic annotation; sampling based on the point cloud semantic segmentation dataset to construct a training set and a test set, training a neural network model based on the training set and the test set to perform parameter training and quantitative evaluation, and selecting a neural network model with optimal convergence performance and minimum computational complexity as the deep learning point cloud semantic segmentation model.

[0008] Optionally, the plane detection on the point cloud of the predefined static scene semantic category to obtain a candidate plane set and corresponding normal information of the plane comprises: sequentially traversing each type of static semantic point cloud of the current frame, constructing an octree for the point cloud of each type of semantic, splitting the octree until the size of each leaf node of the octree is smaller than a preset cluster threshold to obtain a split octree; iteratively performing plane estimation test from any leaf node of the split octree to perform plane fitting on the current point cloud cluster, and if the size of outliers of the current point cloud cluster is greater than a preset percentage of the point cloud, it is determined that the current point cloud cluster does not belong to a plane, otherwise it is determined that the plane condition is met; wherein if all leaf nodes of the current node do not pass the above determination, the current node becomes a new leaf node, and all leaf nodes of the current node are added to the point cloud cluster of the nearest upper root node of the current node, and the new point cloud cluster is subjected to plane estimation test, and if the current node passes the plane estimation test, it enters the merging stage, otherwise it continues to merge upwards; when the plane grows, a neighbor graph is constructed based on the point cloud cluster that meets the plane condition, wherein the vertices of the neighbor graph are point clouds, and the edges of the neighbor graph are edges that meet the neighbor condition when performing k-nearest neighbor search on any point cloud, and the search is performed by a breadth-first algorithm, and the point clouds that do not belong to other planes and meet the distance threshold and the normal variance threshold conditions of the current plane are merged into the current plane; for two planes belonging to the same octree node, if a preset merging condition is met, the two planes are merged, and the merged point cloud cluster is subjected to plane fitting and normal estimation again until no point cloud cluster is updated, to obtain the candidate plane set and the normal information of the plane estimation.

[0009] Optionally, the preset merging condition comprises: there is at least one associated edge between the point clouds contained in the two planes in the neighbor graph; the normal estimated based on the point cloud cluster of the two merged planes is less than a preset value; and there is at least one point cloud sample in any one of the two planes, which meets the plane distance threshold and the normal variance threshold of the estimation of the other plane.

[0010] Optionally, the constructing of the adjacency matrix of the candidate planes in the candidate plane set through the parallel plane detection condition comprises: traversing the plane pairs in the candidate plane set; calculating the distance of the normal approximation of the plane pairs, and if the distance is less than a predetermined threshold, supplementing a first value at the corresponding position of the adjacency matrix, otherwise supplementing a second value; and generating the adjacency matrix based on the traversal result.

[0011] Optionally, the detecting of the subgraph belonging to the coplanar plane by using the adjacency matrix and the clique search algorithm comprises: constructing an adjacency graph according to the adjacency matrix, and performing the clique search algorithm on the adjacency graph to obtain all subgraphs with a size greater than a preset threshold.

[0012] Optionally, the formula for calculating the normal entropy average is:

[0013]

[0014] where i represents the i-th point cloud, k represents the total point cloud combing, n i represents the normal of plane i, and e represents a natural constant.

[0015] The second aspect embodiment of the present application provides a SLAM map topology evaluation device, comprising: an acquisition module configured to acquire point cloud data of a target area; a detection module configured to input the point cloud data into a trained deep learning point cloud semantic segmentation model, output semantic label information of the point cloud, determine a semantic category of the point cloud based on the semantic label information, and perform plane detection on point clouds of a pre-defined static scene semantic category to obtain a candidate plane set and corresponding estimated normal information; a calculation module configured to construct an adjacency matrix of candidate planes in the candidate plane set through a parallel plane detection condition, detect subgroups belonging to coplanar planes using the adjacency matrix and a group search algorithm, and calculate coplanar point variance based on the subgroups; and an evaluation module configured to perform normal condensation clustering on the normal information to obtain a plurality of normal sub-clusters, calculate normal entropy based on the normal of each sub-cluster, calculate a normal entropy average of all sub-clusters, and determine a map topology evaluation result of a simultaneous localization and mapping (SLAM) map of the target area according to the coplanar point variance and / or the normal entropy average.

[0016] The third aspect embodiment of the present application provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the SLAM map topology evaluation method as described in the above embodiments.

[0017] The fourth aspect embodiment of the present application provides a computer readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to implement the SLAM map topology evaluation method as described in the above embodiments.

[0018] Therefore, the present application has at least the following beneficial effects:

[0019] The embodiment of the application can evaluate the pose error of SLAM estimation through local point cloud topology information without the aid of additional measuring equipment in the case of GPS shielding in a large indoor environment, which can promote the development of related algorithms for large and complex artificial environments; by introducing a point cloud semantic segmentation model based on a neural network, accurate semantic information acquisition of complex scenes is realized, and filtering of unstable and target categories not containing potential planes is performed based on semantic labels, thereby reducing the complexity of subsequent calculations; through robust statistical plane detection, plane detection and normal estimation of stable semantic point clouds of each category are performed, which can improve the detection accuracy of candidate planes and the normal estimation accuracy of corresponding planes; based on the principle of SLAM pose estimation, corresponding quantitative evaluation indexes are designed for rotation noise error and translation noise error, and quantitative evaluation is realized for different pose noise components for the first time; by using the coplanar variance statistical calculation method and the average normal entropy statistical calculation method based on normal clustering, the evaluation of SLAM pose is realized, and the like. Thus, the technical problems in the related art that the SLAM pose estimation is limited by the environment and the measuring device, and the precision estimation method is limited, and the like are solved.

[0020] Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following and / or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0021] The above and / or additional aspects and advantages of the application will become apparent and be readily appreciated from the following description, including the accompanying drawings, in which:

[0022] Figure 1 A flowchart of a SLAM map topology evaluation method according to an embodiment of the application;

[0023] Figure 2 A schematic diagram of the influence of rotation noise and translation noise on SLAM pose estimation according to an embodiment of the application;

[0024] Figure 3 A flowchart of a SLAM map topology evaluation method according to an embodiment of the application;

[0025] Figure 4 A schematic diagram of a simulation environment and a mobile robot and radar perception scheme according to an embodiment of the application;

[0026] Figure 5 A schematic diagram of a manually annotated semantic superimposed map visualization according to an embodiment of the application;

[0027] Figure 6 A schematic diagram of plane detection and normal estimation results according to an embodiment of the application;

[0028] Figure 7 A schematic diagram of the binary correlation between the topological metric and the RPE in the case of adding only rotational noise according to an embodiment of the present application is provided;

[0029] Figure 8 A schematic diagram of the binary correlation between the topological metric and the RPE in the case of adding only translational noise according to an embodiment of the present application is provided;

[0030] Figure 9 A schematic diagram of the ternary correlation between the topological metric and the RPE in the case of adding rotational and translational noise according to an embodiment of the present application is provided;

[0031] Figure 10 An example diagram of a SLAM map topological evaluation device according to an embodiment of the present application is provided;

[0032] Figure 11 A structural schematic diagram of an electronic device according to an embodiment of the present application is provided. DETAILED DESCRIPTION

[0033] The embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0034] The SLAM map topological evaluation method, device, electronic device and storage medium of the embodiments of the present application are described below with reference to the accompanying drawings. In view of the problem that the current SLAM development mainly relies on relative pose estimation and absolute pose estimation based on reference estimation metric, but is limited by the environment and measurement device, resulting in difficulty in actual deployment and limited precision estimation method in the above background art, the present application provides a SLAM map topological evaluation method. In the method, the noise existing in the pose estimation is evaluated by the point cloud topological information of the SLAM local map, and the plane detection is optimized by means of deep learning point cloud semantic segmentation, and self-evaluation is performed without the aid of GPS or other measurement equipment. Thus, the problems of SLAM pose estimation being limited by the environment and measurement device, and the limited precision estimation method in the related art are solved.

[0035] Before the SLAM map topology evaluation method of the embodiments of the present application is described, the related art is briefly described, in particular: In recent years, some research attempts to analyze the pose estimation indirectly by using the map topology constructed by the SLAM algorithm, such methods are called topology analysis methods based on SLAM map features, but currently it is in the initial stage. Early methods mainly evaluate based on the topology rules of global point cloud entropy mean, potential plane variance and mutual orthogonal plane, but are affected by the plane detection accuracy, normal estimation error, and the inability to handle the dynamic semantics of complex dynamic time-varying scenes.

[0036] In particular, Figure 1 A flowchart of a SLAM map topology evaluation method provided by an embodiment of the present application.

[0037] As Figure 1 shown, the SLAM map topology evaluation method includes the following steps:

[0038] In step S101, point cloud data of a target area is obtained.

[0039] The target area can be a parking lot, an office, etc. In the embodiments of the present application, for the convenience of description, a three-dimensional simulation environment constructed by a local target indoor scene is taken as an example, including a map containing regular semantic information and a mobile robot and corresponding laser radar configuration.

[0040] In particular, the embodiments of the present application can control the mobile robot to move randomly in the target area, collect point cloud data (Point Cloud=[x,y,z] i ), and construct a point cloud data set and the corresponding pose (Pose) sequence through the data collected by traversing different scenes. It should be noted that when constructing the point cloud SLAM data set, in order to obtain the pose information with controllable noise distribution, different distribution of Gaussian white noise is added to the translation component (t) and the rotation component (R) of the pose information, to obtain the pose with noise (Pose=T est,i ) to simulate the noise fluctuation range estimated by the controllable SLAM algorithm, for subsequent correlation analysis and proof.

[0041] In step S102, the point cloud data is input into the trained deep learning point cloud semantic segmentation model, and the semantic label information of the point cloud is output. Based on the semantic label information, the semantic category of the point cloud is determined, and the point cloud of the pre-defined static scene semantic category is subjected to plane detection to obtain a candidate plane set and the corresponding estimated normal information.

[0042] The semantic label (Label=l i ) of the point cloud is obtained by frame-by-frame labeling of the semantics of the point cloud based on a point cloud visualization semantic labeling tool.

[0043] It can be understood that the embodiments of the present application can input the point cloud data into the trained deep learning point cloud semantic segmentation model, output the semantic label information of the point cloud (with certain error), determine the semantic category of the point cloud, process the predefined static scene semantic category based on the plane detection algorithm, and obtain the candidate plane set and the normal information of the corresponding plane.

[0044] It should be noted that the embodiments of the present application perform sliding window sampling on the continuous point cloud in the point cloud SLAM data set, obtain the point cloud in the current window, and sequentially input each frame of point cloud into the trained final semantic segmentation model, so as to obtain the model semantic segmentation label of each frame of point cloud. Before performing plane detection and normal estimation on each type of static semantic point cloud, each frame of point cloud is sequentially input into the trained semantic segmentation model, and the unstable features in the scene are removed based on the predefined label category, filtering potential moving targets, non-stable targets and target semantic point clouds not containing planes, such as dynamic and potential dynamic irregular targets, and only retaining stable point cloud semantics such as walls, floors, ceilings and weighing columns. Plane segmentation is performed on each type of point cloud remaining.

[0045] In the embodiments of the present application, before the point cloud data is input into the trained deep learning point cloud semantic segmentation model, it further includes: collecting point cloud data of an environment; constructing a point cloud semantic segmentation data set based on the point cloud data manually annotated with semantic labels; sampling to construct a training set and a test set based on the point cloud semantic segmentation data set, and performing parameter training and quantitative evaluation of the neural network model using the training set and the test set respectively, and taking the neural network model with the optimal convergence performance and the minimum calculation complexity as the deep learning point cloud semantic segmentation model.

[0046] The point cloud data of the environment can be obtained by a manually controlled robot randomly walking in a simulation environment.

[0047] Specifically, the embodiments of the present application construct a model trained based on point cloud semantic segmentation data, including: collecting point cloud data of an environment by a robot carrying a laser radar sensor, annotating typical indoor scene semantics based on manual point cloud semantic annotation, i.e., constructing a point cloud semantic segmentation data set based on radar point cloud and semantic labels, then training a neural network model, sampling to construct a training set and a test set based on the constructed point cloud semantic segmentation data set, calculating the error of the model output result and the semantic label based on the segmentation accuracy, driving the model parameter adjustment through error back propagation and stochastic gradient descent, and ensuring that the model converges through enough training rounds, and selecting a lightweight model suitable for deployment of a robot mobile platform as the final model by comprehensively comparing the model segmentation accuracy, calculation speed, calculation complexity, and overall model parameter size.

[0048] In the embodiment of the present application, the point cloud of the predefined static scene semantic category is subjected to plane detection to obtain a candidate plane set and corresponding estimated normal information, including: traversing each type of static semantic point cloud of the current frame in turn, constructing an octree for the point cloud of each type of semantic, splitting the octree until the size of each leaf node of the octree is less than a preset cluster threshold to obtain a split octree; performing iterative plane estimation test from any leaf node of the split octree to perform plane fitting on the current point cloud cluster, if the size of outliers of the current point cloud cluster is greater than a preset percentage of the point cloud, it is determined that the current point cloud cluster does not belong to a plane, otherwise it is determined that the plane condition is met; wherein, if all leaf nodes of the current node do not pass the above determination, the current node becomes a new leaf node, and all leaf nodes of the current node are added to the point cloud cluster of the nearest upper root node of the current node, and the new point cloud cluster is subjected to plane estimation test, if the current node passes the plane estimation test, it enters the merging stage, otherwise it continues to merge upwards; when the plane grows, a neighbor graph is constructed based on the point cloud clusters that meet the plane condition, wherein the vertices of the neighbor graph are point clouds, and the edges of the neighbor graph are edges that meet the neighbor condition when performing k-nearest neighbor search on any point cloud, and the neighbor graph is searched by a breadth-first algorithm, and the point clouds that do not belong to other planes and meet the distance threshold and the normal variance threshold conditions of the current plane are merged into the current plane; for two planes belonging to the same octree node, if a preset merging condition is met, the two planes are merged, and the merged point cloud cluster is subjected to plane fitting and normal estimation again until no point cloud cluster is updated, to obtain a candidate plane set and corresponding normal information of the plane.

[0049] The preset merging condition includes: there is at least one associated edge between the point clouds contained in the two planes in the neighbor graph; the normal estimated based on the point cloud clusters of the two merged planes is less than a preset value; there is at least one point cloud sample in any one of the two planes, which meets the plane distance threshold and the normal variance threshold of the estimation of the other plane.

[0050] It can be understood that the specific steps of the plane detection in the embodiment of the present application are:

[0051] Firstly, each type of static semantic point cloud of the current frame is traversed in turn, an octree is constructed for the point cloud of each type of semantic belonging to the current frame, the octree is split based on space until the size of each leaf node is smaller than a preset cluster threshold; iterative plane estimation test is performed from the octree leaf node, if the current node passes the plane test, it enters the merging stage, if the eight leaf nodes of the current node all fail the test, the current node becomes a new leaf node, and the previous eight leaf nodes are added to its cluster, and plane estimation test is performed on the new cluster. The plane estimation test is to perform RANSAC-based plane fitting on the current cluster, and when the size of the outlier points is greater than a preset percentage (such as 25%), the current point cloud cluster does not belong to a plane, otherwise it satisfies the plane condition.

[0052] When the plane grows, the point cluster that satisfies the plane condition is subjected to construction of a near-neighbor graph G, wherein the vertices of G are point clouds, and the edges of G are edges that satisfy the near-neighbor condition when performing k-nearest neighbor search on any point, and the breadth-first algorithm is used for searching, and points that do not belong to other planes and satisfy the distance threshold and the normal direction threshold conditions to the current plane are merged into the current plane.

[0053] Then, plane merging is performed, for two planes A and B belonging to the same octree node, when the following conditions are satisfied: (1) for the near-neighbor graph G, there is at least one associated edge between the points contained in A and B (that is, when kNN search is performed, the distance threshold is satisfied, and here the graph G above is reused to simplify the calculation process); (2) the normal estimated based on the merged A and B point cloud clusters is less than max(N A , N B ); (3) there is at least one point cloud sample in A that satisfies the distance threshold and the normal variance threshold of the plane estimated by B, and vice versa. A and B can be merged, and plane fitting and normal estimation are performed again on the merged point cluster.

[0054] Through iteration of the above steps until no cluster is updated, the obtained plane set is the final result of plane detection, and all point cloud plane clusters estimated from the point clouds of all frames in the current window are added to a set, which contains all planes and corresponding normal information.

[0055] In step S103, an adjacency matrix of the candidate planes in the candidate plane set is constructed by parallel plane detection conditions, a subgraph belonging to a coplanar plane is detected by using the adjacency matrix and a clique search algorithm, and coplanar point variance is calculated based on the subgraph.

[0056] It can be understood that, based on the candidate plane set, the parallel plane detection based on the cluster search is performed, a plurality of parallel plane sub-sets are obtained, and then the co-plane set detection is further performed based on the approximate co-plane condition, points belonging to the co-plane are taken as a group, the point cloud of the group is calculated, and the co-plane variance (CPV) is obtained.

[0057] In the embodiment of the application, the adjacency matrix of the candidate planes in the candidate plane set is constructed through the parallel plane detection condition, including: traversing the plane pairs in the candidate plane set; calculating the distance of the normal approximation of the plane pairs, if the distance is less than a predetermined threshold, a first value is supplemented in the corresponding position of the adjacency matrix, otherwise a second value is supplemented; and generating the adjacency matrix based on the traversal result.

[0058] Specifically, for the plane set N in the set, a N*N all-0 adjacency matrix is first constructed, and the two plane pairs P1 and P2 in the set are traversed, for the normal approximation of the plane pairs, it is assumed that the two planes are parallel (A1=A2, B1=B2, C1=C2), and the distance d of the two planes is calculated, for the plane pairs with a distance d less than a predetermined threshold, an edge 1 is added to the corresponding position of the adjacency matrix, otherwise 0 is filled.

[0059] P1: A1x+B1y+C1z+D1=0,

[0060] P2: A2x+B2y+C2z+D2=0,

[0061]

[0062] In the embodiment of the application, the sub-clusters belonging to the co-planes are detected by using the adjacency matrix and the cluster search algorithm, including: constructing an adjacency graph according to the adjacency matrix, and performing the cluster search algorithm on the adjacency graph to obtain all sub-clusters with a size greater than a preset threshold.

[0063] The preset threshold can be set according to specific conditions, such as 2 or 3.

[0064] Taking the preset threshold of 2 as an example, the adjacency graph D can be constructed based on the adjacency matrix generated based on the traversal result, the cluster search algorithm is performed on D, and all clusters with a size greater than 2 are obtained, that is, a sub-set containing at least 2 co-planes, the point cloud in each sub-set is merged, and the plane variance of the merged co-plane points is calculated, that is, the minimum eigenvalue of the corresponding point cloud covariance matrix.

[0065] In step S104, the normal information is subjected to normal condensation clustering to obtain a plurality of normal sub-clusters, the normal of each sub-cluster is calculated for normal entropy, the normal entropy average of all sub-clusters is calculated, and the map topology evaluation result of the simultaneous localization and mapping (SLAM) map of the target region is determined according to the coplanar point variance and / or the normal entropy average.

[0066] wherein the calculation formula of the normal entropy average is:

[0067]

[0068] wherein i represents the i-th point cloud, k represents the total point cloud combing, n_i represents the normal of plane i, and e represents a natural constant. It can be understood that the embodiments of the present application can perform normal condensation clustering on the normal information, calculate the normal entropy based on the normal of each cluster, and determine the map topology evaluation result of the simultaneous localization and mapping (SLAM) map of the target region according to the coplanar point variance and / or the normal entropy average.

[0069] It can be understood that the embodiments of the present application adopt condensation clustering for the normal value of each plane in the set, obtain a plurality of normal approximate sub-clusters, and then calculate the normal entropy average (Plane Normal Entropy, PNE) of all clusters based on the calculation formula for each cluster with a size greater than 3 (k>3), n i is the normal of plane i. Wherein the coplanar variance reflects the translation noise component in the relative pose error, the larger the value is, the larger the translation noise (noise of t) component is, and the normal entropy average reflects the correlation of the rotation error (noise of R) noise component, the larger the value is, the larger the rotation noise component is.

[0070] In the embodiments of the present application, the point cloud SLAM is used to estimate the pose change of the mobile robot, and the principle is to estimate the relative motion of the robot based on the configuration algorithm of the two consecutive frames of point clouds, mainly including three-dimensional translation motion and rotation motion. Wherein the translation motion can be represented as a t vector, t=[x, y, z] is the relative displacement of the robot along the world three-dimensional coordinate system in a time slice, and the rotation matrix R of the robot around the X-Y-Z axis is represented, and the total pose matrix is defined as:

[0071]

[0072]

[0073] However, due to the existence of sensor measurement noise, the rigid body motion estimation of the robot based on the two consecutive frames of point clouds, i.e. calculating T est , will inevitably introduce errors. The accurate result obtained based on the measurement device or GPS is defined as T gtThe relative pose error (RPE) between two adjacent frames at time i and i+Δt can be defined as

[0074]

[0075] However, when the GPS positioning information is unavailable and the motion of the robot exceeds the measurement range of the laser tracker and other devices, the corresponding T gt , resulting in the above RPE being unable to be calculated.

[0076] Therefore, it is also necessary to consider replacing the existing RPE estimation scheme with the topological consistency measurement introduced by the above process, and to verify the correlation of the proposed map topological measurement and RPE with the absolute accurate pose information provided by the simulation environment.

[0077] Specifically, the SLAM pose estimation is shown in the above formula, where the rotation component R can be described as a 3x3 rotation matrix, which includes the rotation change of the robot motion along the X-Y-Z axis; the translation component t can be described as a 3x1 vector, which includes the translation change of the robot motion along the X-Y-Z axis. As shown in Figure 2 , the rotation component noise will cause the change of the potential plane normal in the scene, thereby causing the randomness of the overall normal to increase; although it will also cause the change of the plane point cloud variance, the fluctuation is difficult to estimate. The noise of the translation component will cause the translation change of the originally registered plane along the X / Y / Z axis, so that the originally single plane fitting point will be jittered, which may generate multiple planes, thereby causing the estimated plane variance to increase. Based on the above phenomenon, the coplanar point cloud variance estimation and the average normal entropy estimation based on the normal clustering are proposed to estimate the translation component and the normal component noise of the pose noise, and the correlation verification is performed through the simulation data.

[0078] In summary, the SLAM map topological evaluation method of the embodiments of the present application, the flow is shown in Figure 3 , comprising:

[0079] (1) Point cloud semantic segmentation data construction;

[0080] (2) Neural network point cloud semantic segmentation model training;

[0081] (3) Point cloud SLAM data set construction;

[0082] (4) Real-time point cloud semantics acquisition;

[0083] (5) Robust plane detection and normal estimation;

[0084] (6) Coplanar point cloud variance estimation;

[0085] (7) Estimation of average normal entropy based on normal clustering;

[0086] (8) Correlation analysis verification.

[0087] The SLAM map topology evaluation method is illustrated below with a specific example:

[0088] Among them, mobile robot experimental platforms such as Figure 4 As shown, it includes a low-cost LiDAR, display, and computing unit mounted on top of the robot, as well as the mobile robot chassis. The simulation environment for the test deployment is as follows: Figure 4 As shown, it includes two large underground parking garages and a medium-sized office, each with the same scale as the local scene, as well as two open-source square office scenes. A simulation platform consistent with the radar of the robot prototype system is built in the simulation environment.

[0089] Step 1: Point Cloud Semantic Segmentation Data Construction. A robot, manually controlled, randomly moves and traverses four simulated scenarios multiple times, acquiring radar point cloud data at 5 / 10Hz. Then, point cloud semantic annotation is performed manually to obtain semantic label information for each frame of the point cloud. Visualization comparison based on the annotated point clouds is shown below. Figure 5 As shown, the upper part of each graph is a visualization result colored by point cloud height, and the lower part is a visualization result colored by label. The current scene annotation categories include five categories of scene semantics: ground, wall, load-bearing column, table, and chair.

[0090] Step 2: Training the Neural Network Point Cloud Semantic Segmentation Model. The point cloud semantic segmentation dataset collected in Step 1 was randomly divided into two subsets: 60% and 40%. Several neural network models were selected for training and evaluation in each subset, and the model parameters were converged using label constraints. The local testing hardware environment consisted of a GPU server, including an Intel i9-12900K CPU, 128GB of RAM, an NVIDIA RTX 3090@24GB graphics card, and a large-capacity hard drive. The segmentation accuracy of the model was verified through actual testing, and the model with the best overall performance was selected as the candidate ensemble model. The results are shown in Table 1. Three typical neural network models, MinkowskiNet, CylinderNet, and SPVCNN, were selected for testing. Among them, MinkowskiNet achieved the best average intersection-over-union (IoU) and class-based IoU, and also had the highest speed on the current testing hardware platform. Therefore, this model was selected as the final semantic segmentation model for ensemble.

[0091] Table 1

[0092]

[0093] where C is the semantic class of the scene, TP, FP and FN are the number of true positive, false positive and false negative of the corresponding class respectively, and the class-based intersection over union does not need to be accumulated and averaged.

[0094]

[0095] Step 3: Point cloud SLAM dataset construction. By manually controlling the robot to move randomly in the opposite scene, the radar point cloud data and the robot pose matrix T at the corresponding time are collected at 5 / 10 Hz to construct a SLAM dataset containing continuous point clouds and poses.

[0096] Step 4: Real-time point cloud semantic acquisition. Based on the sliding window (window length 5, step 5, i.e. non-overlapping sampling), the continuous point cloud is sampled, and each frame of point cloud is input into the MinkowskiNet trained in step 2 for processing to obtain the segmentation label. Among them, the non-static target (i.e. table and chair) is filtered, and only the wall, ground and load-bearing column point clouds that can feedback the stable topological information of the scene are retained.

[0097] Step 5: Robust statistical plane detection and normal estimation. The robust statistical plane detection algorithm is used to detect the plane of each type of stable semantic point cloud in each frame of point cloud, and the obtained plane point cloud and normal information are sequentially added to the candidate plane set. The entire set contains all potential planes in the current sliding window, as shown in Figure 6 .

[0098] Step 6: Co-planar point cloud variance estimation. Based on the current candidate plane set, an adjacency matrix is constructed and initialized by filling all 0. Each two plane pairs [P1, P2] are traversed in turn, and the normal difference |N1, N2|>1-∈, where ∈ is a minimum threshold, is determined as parallel planes. At this time, it can be assumed that the A, B and C of the two plane formulas are consistent, the distance d between the two planes is calculated based on the calculation formula, and d<3∈ (the multiple is set based on the situation) is determined. For the plane pairs that meet the above co-planar conditions, 1 is added to the corresponding positions of the adjacency matrix (including the two blocks symmetric to the main diagonal), representing an edge.

[0099] Based on the final adjacency matrix, an adjacency graph is generated, the vertices of the graph are the candidate planes, and the edges of the graph represent the plane pairs connected by the two vertices as co-planar. Then based on the group search, all groups with a group size greater than 2 are obtained, i.e. each group contains the corresponding co-planar.

[0100] Merge the planar point cloud contained in each group, and calculate the eigenvalue corresponding to the covariance matrix of the current point set, where the minimum eigenvalue is the variance of the planar point cloud. Sort the planar point cloud variances of all groups, and the median is the average coplanar variance estimation (CPV).

[0101] Step 7: Average plane normal entropy estimation (PNE) based on normal clustering. Based on all the normals of the current candidate plane set, perform agglomerative clustering, initially taking the normal of each plane as an independent cluster, and merging based on a preset threshold with adjacent normals, and stopping based on a preset condition. A number of clusters based on normal approximation are obtained, and for each cluster, the entropy based on the normal is independently calculated, and then the average of the normal entropies of all clusters (such as formula 7) is the result.

[0102] Step 8: In order to further verify the correctness of the two SLAM algorithm-based local point cloud map topology information pose noise estimation designed, common binary correlation measures Pearson, Spearman and Kendall, and multi-relation ternary correlation measure MultiRelation (MR) are introduced, and the calculation formula is as follows. Where n observations of variables Y i ,=1,…, can construct a k*n matrix Y. Since each k may come from different observation backgrounds, such as RPE, CPV and PNE in this embodiment, normalization is needed by row to obtain the corresponding standardized matrix S. Then the sample correlation matrix R = SS T , where the minimum eigenvalue of R is λ(R).

[0103] MR(Y1,…,Y k )=1-λ(R)=1-λ(SS T )

[0104] In order to control the noise distribution and scale, rotation noise, translation noise and rotation-translation noise are added to the poses in the sliding window with a length of 5, respectively. Among them, the rotation noise takes [-pi / 180,pi / 180]*Scale uniform distribution noise around the Z axis yaw, and [-0.05*pi / 180,0.05*pi / 180]*Scale uniform distribution noise around the XY axis roll and pitch; The translation noise takes [-0.1,0.1]*Scale uniform distribution noise on the XY axis, and [-0.005,0.005]*Scale uniform distribution noise on the Z axis. Mainly considering that the robot moves mainly along the 2D plane and rotates in the indoor motion. While Scale is configured as [1,1.5,2] three scales. And compared with the early developed MME, MPV and MOM three measures, the above three measures do not distinguish the independent correlation analysis of the translation component and the rotation component.

[0105] wherein, as shown in Figure 7 Among the binary correlation of topological metrics with RPE with only added rotational noise, the correlation of planar normal entropy (PNE) metric with relative pose error (RPE) performs the highest under three different scales of noise, which indicates that planar PNE can correctly feedback the relative pose error caused by rotational noise.

[0106] As shown in Figure 8 Among the binary correlation of topological metrics with RPE with only added translational noise, the correlation of coplanar point cloud variance (CPV) metric with relative pose error (RPE) performs the highest under three different scales of noise, which indicates that planar CPV can correctly feedback the relative pose error caused by translational noise.

[0107] As shown in Figure 9 Among the ternary correlation of topological metrics with RPE with added translational and rotational noise, the ternary correlation of planar normal entropy (PNE) metric, coplanar point cloud variance (CPV) metric and relative pose error (RPE) performs the highest under three different scales of noise, which indicates that the integration of two independent metrics has the possibility to correctly feedback the relative pose error caused by translational and rotational noise. k ) = 1 - λ(R) = 1 - λ(SS T ) calculation, the ternary correlation of planar normal entropy (PNE) metric, coplanar point cloud variance (CPV) metric and relative pose error (RPE) performs the highest under three different scales of noise, which indicates that the integration of two independent metrics has the possibility to correctly feedback the relative pose error caused by translational and rotational noise.

[0108] Through the binary and ternary correlation metrics, it is verified that the mean of coplanar point cloud variance and planar normal entropy has a stable correlation with the relative pose estimation metric based on the benchmark measurement, which has the possibility to rely only on map topological information as a substitute for SLAM pose estimation in the case of lack of GPS and high-precision measurement devices in indoor environment.

[0109] According to the SLAM map topology evaluation method provided in the embodiments of the present application, the pose error estimated by SLAM can be evaluated through local point cloud topology information without the aid of additional measuring equipment in the case of GPS shielding in a large indoor environment, which can promote the development of related algorithms for large and complex artificial environments; by introducing a point cloud semantic segmentation model based on a neural network, accurate semantic information of a complex scene can be obtained, and filtering of target categories that are unstable and do not contain potential planes is performed based on semantic labels, thereby reducing the complexity of subsequent calculation; through robust statistical plane detection, stable semantic point clouds of each category are subjected to plane detection and normal evaluation, which can improve the detection accuracy of candidate planes and the normal estimation accuracy of corresponding planes; based on the pose estimation principle of SLAM, corresponding quantitative evaluation indexes are designed for rotational noise error and translational noise error, and quantitative evaluation is realized for different pose noise components for the first time; by using a coplanar variance statistical calculation method and a mean normal entropy statistical calculation method based on normal clustering, evaluation of the SLAM pose is realized.

[0110] Secondly, the SLAM map topology evaluation device provided in the embodiments of the present application is described with reference to the accompanying drawings.

[0111] Figure 10 is a block schematic diagram of the SLAM map topology evaluation device in the embodiments of the present application.

[0112] As shown in Figure 10 , the SLAM map topology evaluation device 10 comprises an acquisition module 100, a detection module 200, a calculation module 300 and an evaluation module 400.

[0113] The acquisition module 100 is configured to acquire point cloud data of a target region; the detection module 200 is configured to input the point cloud data into a trained deep learning point cloud semantic segmentation model, output semantic label information of the point cloud, determine semantic categories of the point cloud based on the semantic label information, and perform plane detection on point clouds of predefined static scene semantic categories to obtain a candidate plane set and corresponding estimated normal information; the calculation module 300 is configured to construct an adjacency matrix of candidate planes in the candidate plane set through parallel plane detection conditions, detect subgroups belonging to coplanar planes by using the adjacency matrix and a group search algorithm, and calculate coplanar point variances based on the subgroups; and the evaluation module 400 is configured to perform normal agglomerative clustering on the normal information to obtain a plurality of normal sub-clusters, calculate normal entropy of each sub-cluster, calculate a mean value of normal entropy of all sub-clusters, and determine a map topology evaluation result of a SLAM map of the target region according to the coplanar point variances and / or the mean value of normal entropy.

[0114] It should be noted that the foregoing explanation and description of the SLAM map topology evaluation method embodiments are also applicable to the SLAM map topology evaluation device of this embodiment, which will not be described here again.

[0115] The SLAM map topology evaluation device provided by the embodiment of the application can evaluate the pose error estimated by SLAM through local point cloud topology information without the aid of additional measuring equipment in the case of GPS shielding in a large indoor environment, can promote the development of related algorithms for large and complex artificial environments, and can realize accurate semantic information acquisition of complex scenes through the introduction of a point cloud semantic segmentation model based on a neural network, filter target categories that are unstable and do not contain potential planes based on semantic labels, and reduce the complexity of subsequent calculations. The detection accuracy of candidate planes and the normal estimation accuracy of corresponding planes can be improved through plane detection and normal evaluation of stable semantic point clouds of each category by robust statistical plane detection. Based on the pose estimation principle of SLAM, corresponding quantitative evaluation indexes are designed for rotational noise error and translational noise error, and quantitative evaluation is realized for different pose noise components for the first time. The translational component noise estimation of SLAM pose, the rotational component noise estimation of SLAM pose, and the error estimation of the overall pose are realized by using the coplanar variance statistical calculation method and the average normal entropy statistical calculation method based on normal clustering.

[0116] Figure 11 The electronic device provided by the embodiment of the application is shown in the structural schematic diagram. The electronic device can include:

[0117] The memory 1101, the processor 1102, and the computer program stored in the memory 1101 and executable on the processor 1102.

[0118] The processor 1102 implements the SLAM map topology evaluation method provided in the above embodiment when executing the program.

[0119] Further, the electronic device further includes:

[0120] The communication interface 1103 is used for communication between the memory 1101 and the processor 1102.

[0121] The memory 1101 is used to store the computer program executable on the processor 1102.

[0122] The memory 1101 can include a high-speed RAM (Random Access Memory, Random Access Memory) memory, and can also include a non-volatile memory, such as at least one disk memory.

[0123] If the memory 1101, the processor 1102 and the communication interface 1103 are implemented independently, the communication interface 1103, the memory 1101 and the processor 1102 can be connected with each other through a bus and complete communication between each other. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 11 Only one thick line is used in the figure to represent the bus, but it does not mean that there is only one bus or only one type of bus.

[0124] Optionally, in a specific implementation, if the memory 1101, the processor 1102 and the communication interface 1103 are integrated on a chip, the memory 1101, the processor 1102 and the communication interface 1103 can complete communication between each other through an internal interface.

[0125] The processor 1102 can be a CPU (Central Processing Unit), or an ASIC (Application Specific Integrated Circuit), or an integrated circuit configured to implement one or more embodiments of the present application.

[0126] The embodiment of the present application also provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the SLAM map topology evaluation method.

[0127] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example" or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms is not necessarily for the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0128] In addition, the terms "first", "second", etc. are used only for descriptive purposes and do not connote or imply relative importance or a quantity of the indicated technical features. Thus, a feature defined with "first", "second", etc. can include at least one of the feature, explicitly or implicitly. In the description of the present application, the meaning of "N" is at least two, such as two, three, etc., unless otherwise explicitly and specifically limited.

[0129] Any process or method descriptions or blocks in flow charts or otherwise described herein can be understood as representing modules, segments, or portions of code which include one or more executable instructions for implementing specific logical functions or steps in the process, and alternate implementations are possible. In some embodiments, the processes or methods described in this application can be tailored or adapted by, for example, removing and / or adding one or more of the process blocks described herein or described in other implementations, or changing the order of the process blocks described herein or described in other implementations. The scope of preferred embodiments of this application encompasses these and other alternatives.

[0130] It should be understood that portions of the present application can be realized with hardware, software, firmware or a combination thereof. In the above embodiments, the N steps or methods can be realized with software or firmware stored in a memory and executed by a suitable instruction execution system. As in another embodiment, if realized with hardware, any one or a combination of the following technologies known in the art can be used: discrete logic circuit with logic gates for implementing logic functions on data signals, application specific integrated circuit with suitable combination logic gates, programmable gate array, field programmable gate array, etc.

[0131] Those skilled in the art of the present technology can understand that all or part of the steps carried out by the above-mentioned embodiment methods can be completed by programs instructing related hardware, and the programs can be stored in a computer readable storage medium. When the program is executed, it includes one of the steps of the method embodiment or a combination thereof.

[0132] Although the embodiments of the present application have been shown and described above, it should be understood that the above-mentioned embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-mentioned embodiments within the scope of the present application.

Claims

1. A method of SLAM map topology evaluation, the method comprising: The method comprises the following steps: Obtaining point cloud data of a target region; Inputting the point cloud data into a trained deep learning point cloud semantic segmentation model to output semantic label information of the point cloud, determining a semantic category of the point cloud based on the semantic label information, and performing plane detection on point cloud of a pre-defined static scene semantic category to obtain a candidate plane set and normal information of the corresponding plane; The plane detection on the point cloud of the pre-defined static scene semantic category to obtain the candidate plane set and the normal information of the corresponding plane comprises: sequentially traversing each type of static semantic point cloud of a current frame, constructing an octree for the point cloud of each type of semantic, and splitting the octree until the size of each leaf node of the octree is smaller than a preset cluster threshold to obtain a split octree; iteratively performing plane estimation testing from any leaf node of the split octree to perform plane fitting on a current point cloud cluster, and if the size of outliers of the current point cloud cluster is greater than a preset percentage of the point cloud, it is determined that the current point cloud cluster does not belong to a plane, otherwise it is determined that the plane condition is met; if all leaf nodes of the current node do not pass the above determination, the current node becomes a new leaf node, and all leaf nodes of the current node are added to a point cloud cluster of a nearest upper root node of the current node, and plane estimation testing is performed on the new point cloud cluster; if the current node passes the plane estimation testing, a merging stage is entered, otherwise the merging continues; when the plane grows, a neighbor graph is constructed based on the point cloud clusters that meet the plane condition, wherein the vertices of the neighbor graph are point clouds, and the edges of the neighbor graph are edges that meet a neighbor condition when performing k-neighbor searching on any point cloud, and the point clouds that do not belong to other planes and meet the distance threshold and the normal variance threshold conditions of the current plane are merged into the current plane by breadth-first search; for two planes belonging to the same octree node, if a preset merging condition is met, the two planes are merged, and plane fitting and normal estimation are performed on the merged point cloud cluster again until no point cloud cluster is updated, to obtain the candidate plane set and the normal information of the corresponding plane; An adjacency matrix of the candidate planes in the candidate plane set is constructed through parallel plane detection conditions, a subgraph belonging to a coplanar plane is detected by using the adjacency matrix and a group search algorithm, and coplanar point variance is calculated based on the subgraph; A plurality of normal sub-clusters are obtained by performing normal agglomerative clustering on the normal information, a normal entropy is calculated based on the normal of each sub-cluster, a mean value of the normal entropy of all sub-clusters is calculated, and a map topology evaluation result of a simultaneous localization and mapping (SLAM) map of the target region is determined according to the coplanar point variance and the mean value of the normal entropy.

2. The SLAM map topology evaluation method of claim 1, wherein, Before the point cloud data is input into the trained deep learning point cloud semantic segmentation model, the method further comprises: Collecting point cloud data of an environment; Based on artificial point cloud semantic annotation of the point cloud data, a point cloud semantic segmentation data set is constructed; The training set and the test set are constructed by sampling based on the point cloud semantic segmentation dataset, the neural network model is trained and quantitatively evaluated by using the training set and the test set respectively, and the neural network model with optimal convergence performance and minimum calculation complexity is taken as the deep learning point cloud semantic segmentation model.

3. The SLAM map topology evaluation method of claim 1, wherein, The preset merging condition comprises: There is at least one associated edge between the points contained in the two planes in the near-neighbor graph; The normal estimated based on the point cloud cluster of the two merged planes is less than a preset value; There is at least one point cloud sample in any one of the two planes, which satisfies the plane distance threshold and the normal variance threshold estimated based on the other plane.

4. The method of claim 1, wherein, The method for constructing the adjacency matrix of the candidate plane in the candidate plane set through the parallel plane detection condition comprises: Traversing the plane pairs in the candidate plane set; Calculating the distance of the normal approximation of the plane pairs, if the distance is less than a predetermined threshold, a first value is supplemented in the corresponding position of the adjacency matrix, otherwise a second value is supplemented; The adjacency matrix generated based on the traversal result.

5. The SLAM map topology evaluation method of claim 1 or 4, wherein, The method for detecting the subgroups belonging to the coplanar plane based on the adjacency matrix and the group search algorithm comprises: Constructing an adjacency graph according to the adjacency matrix, and performing the group search algorithm on the adjacency graph to obtain all subgroups with a size greater than a preset threshold.

6. The method of claim 1, wherein, The calculation formula of the normal entropy mean value is: ， wherein, denotes the i-th point cloud, denotes the total point cloud, denotes the plane normal to the plane, denotes the natural constant.

7. A SLAM map topology evaluation apparatus, characterized by, The method comprises: An acquisition module configured to acquire point cloud data of a target region; A detection module configured to input the point cloud data into the trained deep learning point cloud semantic segmentation model, output semantic label information of the point cloud, determine a semantic category of the point cloud based on the semantic label information, and perform plane detection on the point cloud of a predefined static scene semantic category to obtain a candidate plane set and normal information of the corresponding plane. The plane detection on the point cloud of the predefined static scene semantic category obtains a candidate plane set and corresponding plane normal information, including: sequentially traversing each type of static semantic point cloud of the current frame, constructing an octree for the point cloud of each type of semantic, splitting the octree until the size of each leaf node of the octree is smaller than a preset cluster threshold, obtaining a split octree; iteratively performing plane estimation testing from any leaf node of the split octree to perform plane fitting on the current point cloud cluster, if the size of outliers of the current point cloud cluster is greater than a preset percentage of the point cloud, it is determined that the current point cloud cluster does not belong to a plane, otherwise it is determined that the plane condition is met; if all leaf nodes of the current node do not pass the above determination, the current node becomes a new leaf node, and all leaf nodes of the current node are added to the point cloud cluster of the nearest upper root node of the current node, and the new point cloud cluster is subjected to plane estimation testing, if the current node passes the plane estimation testing, it enters the merging stage, otherwise it continues to merge upwards; when the plane grows, a neighbor graph is constructed based on the point cloud clusters that meet the plane condition, wherein the vertices of the neighbor graph are point clouds, and the edges of the neighbor graph are edges that meet the neighbor condition when performing k-nearest neighbor search on any point cloud, and the search is performed by a breadth-first algorithm, and the point clouds that do not belong to other planes and meet the distance threshold and normal variance threshold conditions of the current plane are merged into the current plane; for two planes belonging to the same octree node, if a preset merging condition is met, the two planes are merged, and the merged point cloud cluster is subjected to plane fitting and normal estimation again until no point cloud cluster is updated, obtaining the candidate plane set and corresponding plane normal information; The calculation module is configured to construct an adjacency matrix of the candidate planes in the candidate plane set through a parallel plane detection condition, detect subgroups belonging to coplanar planes using the adjacency matrix and a group search algorithm, and calculate coplanar point variance based on the subgroups. The evaluation module is configured to perform normal agglomerative clustering on the normal information to obtain a plurality of normal sub-clusters, calculate normal entropy based on the normal of each sub-cluster, calculate the mean value of the normal entropy of all sub-clusters, and determine a map topology evaluation result of a simultaneous localization and mapping (SLAM) map of a target region according to the coplanar point variance and the mean value of the normal entropy.

8. An electronic device, comprising: The computer program is stored in the memory and executable on the processor, and the processor executes the program to implement the SLAM map topology evaluation method according to any one of claims 1-6. The program is executed by the processor to implement the SLAM map topology evaluation method according to any one of claims 1-6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, ​

Images