A method for dynamic map construction and tile-based differential updates

Abstract

This invention discloses a dynamic map construction and tile-based differentiated update method. The method divides the map into four types of map tiles: stable tiles, slowly changing tiles, low-dynamic tiles, and ground tiles. Different positioning weights and update strategies are assigned to each type of map tile. A change rate detection algorithm dynamically adjusts the map tile type to achieve adaptive map management. This invention improves the robustness of positioning and registration, reduces interference from dynamic objects, realizes classified management and differentiated updates of map tiles, ensures the long-term reliability and accuracy of the map, and supports dynamic loading and efficient updates of infinitely large maps.

Description

Technical Field

[0001] This invention belongs to the field of SLAM technology, specifically relating to the design of a dynamic map construction and tile-based differential update method. Background Technology

[0002] Simultaneous Localization and Mapping (SLAM) is a core technology in mobile robotics, autonomous driving, and drones. In large-scale, long-term operating scenarios, SLAM systems not only need to solve localization problems caused by map changes, but also the reliability issues of long-term map maintenance and dynamic updates. Existing SLAM systems generally suffer from the following problems: (1) Insufficient adaptability to dynamic environment: There are objects with different dynamic characteristics in the scene (ground, buildings, trees, vehicles, etc.). The environment will change over time, introducing incorrect matching during localization, resulting in a decrease in localization accuracy or even errors. In addition, some SLAM systems only perform static / dynamic binary classification and cannot model the change pattern in fine granularity.

[0003] (2) The coupling between localization and mapping is too strong: The residual dynamic or semi-dynamic data in the existing SLAM system map will directly interfere with localization optimization. At the same time, the map update strategy is too simple, which can easily lead to map drift or pollution.

[0004] (3) Limited map size: The existing SLAM system loads the global point cloud map as a whole, which puts a lot of pressure on memory and storage, making it difficult to support ultra-large scale or infinitely expandable scenes.

[0005] Therefore, how to achieve a scalable, updatable, and long-term maintainable dynamic map representation and update mechanism while ensuring localization robustness has become an important research direction for current SLAM technology. Summary of the Invention

[0006] The purpose of this invention is to address the problems of decreased positioning accuracy and deteriorated map quality in existing SLAM systems caused by long-term changes in dynamic objects and the environment. This invention proposes a dynamic map construction and tile-based differentiated update method to improve the robustness of positioning and registration, reduce interference from dynamic objects, realize the classification management and differentiated update of map tiles, ensure the long-term reliability and accuracy of the map, and support dynamic loading and efficient updating of infinitely large maps.

[0007] The technical solution of this invention is: a method for dynamic map construction and tile-based differential updates, comprising the following steps: S1. Divide the global map into multiple map tiles according to a fixed spatial size, and remove highly dynamic targets in real time at the SLAM front end using a moving object detection method.

[0008] S2. Automatically classify each map tile into a stable tile, a slowly changing tile, a low-dynamic tile, or a ground tile.

[0009] S3. Generate a map based on the original laser point cloud.

[0010] S4. Locate and register the generated map during initialization, and assign different weights to different types of map tiles.

[0011] S5. Calculate the rate of change of the registered map tiles and dynamically adjust the map tile type based on the mean and variance of the rate of change over multiple time windows.

[0012] S6. Different update methods are used to update stable blocks, slowly changing blocks, low-dynamic blocks, and ground blocks respectively, so as to achieve differentiated updates for different types of map blocks.

[0013] Furthermore, each map block obtained in step S1 independently stores point cloud data, statistical features, and dynamic state parameters.

[0014] Furthermore, the map tiles obtained in step S1 support on-demand loading and unloading to enable local access.

[0015] Furthermore, the stable blocks in step S2 remain unchanged, the slowly changing blocks change continuously but slowly, the low-dynamic blocks exist within a certain time scale but change significantly, and the ground blocks are obtained by the ground extraction algorithm and are processed as slowly changing blocks in the update and localization process.

[0016] Furthermore, the specific steps of the ground extraction algorithm are as follows: A1. Using the location of the mobile robot as the center, convert the laser point cloud into polar coordinates, and divide the laser point cloud into partitions according to the distance, angle and radius of the laser point cloud relative to the center to obtain point cloud partitions.

[0017] A2. For each point cloud partition, the group of points with the lowest position is selected as candidate ground points. SVD is used to fit the local plane, and the distance from the remaining points in the point cloud partition to the local plane is calculated. If the distance is less than a preset threshold, the corresponding point is classified as a ground point, and the ground segmentation point cloud extracted in the current frame is obtained.

[0018] A3. Transform the ground segmentation point cloud extracted from the previous frame into the current coordinate system, supplement the missing ground points in the current frame through multi-frame fusion, and at the same time perform a weighted average of the existing ground points in the current frame to obtain the ground block result.

[0019] Furthermore, step S3 includes the following sub-steps: S31. Perform ground separation on the original laser point cloud to obtain ground point cloud and non-ground point cloud.

[0020] S32. Store the ground point cloud directly into the ground block, and at the same time, divide the non-ground point cloud into spatial blocks.

[0021] S33. Set the initial type of all map tiles in the first generated map to slowly changing tiles.

[0022] Furthermore, the formula for assigning different weights to different types of map tiles in step S4 is as follows: in Indicates the first i A map tile, Indicates the first i The weight of each map tile, Table of stable block weights, This indicates that the block weights are changing slowly. This indicates low dynamic block weights, and has .

[0023] Furthermore, the overall pose optimization objective function for localization and map registration in step S4 is: in This represents the map and observation laser error function. This indicates the current observation point cloud in the map tile. A subset within the coverage area.

[0024] Furthermore, step S5 includes the following sub-steps: S51. Based on the registered map tiles And the current observation point cloud in the map tile subset within the coverage area The map tiles were calculated. rate of change : in This indicates the number of points or voxels occupied.

[0025] S52. For the same map tile, within the time window T Record the rate of change multiple times and calculate the mean of the multiple rates of change. and variance .

[0026] S53, if If so, the map tile type will be updated to stable tiles. In the case of a certain condition, the map tile type will be updated to low-dynamic tile; otherwise, the map tile type will be updated to slowly changing tile. This represents the empirical threshold for the stable block mean. This represents the empirical threshold for stable block variance. This represents the empirical threshold for the low dynamic block mean. This represents the empirical threshold for low dynamic block variance.

[0027] Furthermore, step S6 specifically includes: For stable blocks, no updates are performed by default; only the rate of change is counted.

[0028] For slowly changing or ground-level patches, voxel downsampling is performed on old map points to make them sparse, and then the new observation point cloud is fused in: ,in This indicates voxel downsampling.

[0029] For low-dynamic blocks, directly replace them with the latest observation point cloud: .

[0030] The beneficial effects of this invention are: (1) The present invention divides the global map into multiple map blocks according to a fixed spatial size. Each map block independently stores point cloud data, statistical features and dynamic status parameters. It supports on-demand loading and unloading to achieve local access. While ensuring that the total size of the map is not limited, it supports fast network transmission and online loading, and is suitable for long-term operation and large-scale environment.

[0031] (2) The present invention divides the map into four types of map tiles and assigns different positioning weights to each type, so that stable structures (such as buildings) are given higher weights and can dominate the positioning optimization process, thus significantly reducing the impact of dynamic residues on positioning.

[0032] (3) This invention dynamically adjusts the map tile type through a change rate detection algorithm. It dynamically adjusts the type based on long-term / multiple map tile change rate statistics, which can not only effectively track map changes, but also is simple and efficient to calculate. When the environment type changes (such as buildings being removed), the variance can also be increased. The stable blocks are gradually degenerated into slowly changing blocks or low-dynamic blocks. Attached Figure Description

[0033] Figure 1 The diagram shown is a flowchart of a dynamic map construction and tile-based differential update method provided by an embodiment of the present invention. Detailed Implementation

[0034] Exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the drawings are merely exemplary and are intended to illustrate the principles and spirit of the invention, and are not intended to limit the scope of the invention.

[0035] This invention provides a method for dynamic map construction and tile-based differential updates, such as... Figure 1 As shown, the process includes the following steps S1 to S6: S1. Divide the global map into multiple map tiles according to fixed spatial dimensions (such as cubes or two-dimensional grid bars), and remove high dynamic targets (pedestrians, vehicles, etc.) in real time at the SLAM front end through moving object detection methods.

[0036] In this embodiment of the invention, each map tile independently stores point cloud data, statistical features, and dynamic status parameters.

[0037] In this embodiment of the invention, map tiles support on-demand loading and unloading to enable local access.

[0038] S2. Automatically classify each map tile into a stable tile, a slowly changing tile, a low-dynamic tile, or a ground tile.

[0039] In this embodiment of the invention, stable blocks such as fixed buildings, walls, and columns remain almost unchanged.

[0040] Slowly changing segments, such as trees, shrubs, and long-term evolutionary structures, change continuously but slowly.

[0041] Low-dynamic segments, such as parked vehicles and mobile devices, exist within a certain time scale but change significantly.

[0042] Ground segments are obtained by ground extraction algorithms and are processed as slowly changing segments during updates and localization.

[0043] In this embodiment of the invention, the specific steps of the ground extraction algorithm are as follows: A1. Convert the laser point cloud to polar coordinates, centered on the location of the mobile robot. r , θ , z The laser point cloud is divided into sections based on its distance (far, medium, near), angle, and radius relative to the center, resulting in point cloud sections (i.e., divided into small patches to handle uneven ground).

[0044] A2. For each point cloud partition, the group of points with the lowest position is selected as candidate ground points. SVD (Singular Value Decomposition) is used to fit a local plane, and the distance *dist* from the remaining points within the point cloud partition to the local plane is calculated. If the distance *dist* is less than a preset threshold, the corresponding point is classified as a ground point, resulting in the ground segmentation point cloud extracted for the current frame. The threshold is adjusted based on the distance between patches and the variance of the point clouds within a patch. After processing all patches, the ground segmentation point cloud extracted for the current frame is obtained.

[0045] A3. Transform the ground segmentation point cloud extracted from the previous frame into the current location coordinate system, and supplement the missing (occluded, etc.) ground points in the current frame through multi-frame fusion (union processing). At the same time, perform weighted averaging on the existing ground points in the current frame to obtain the ground block result.

[0046] In this embodiment of the invention, since the laser point cloud acquisition positions of the two frames are different and the original laser data is in the sensor coordinate system, it is necessary to use the positioning information to transform to the same coordinate system (generally to the laser coordinate system of the current frame).

[0047] S3. Generate a map based on the original laser point cloud.

[0048] Step S3 includes the following sub-steps S31 to S33: S31. Perform ground separation on the original laser point cloud to obtain ground point cloud and non-ground point cloud.

[0049] S32. Store the ground point cloud directly into the ground block, and at the same time, divide the non-ground point cloud into spatial blocks.

[0050] S33. Set the initial type of all map tiles in the first generated map to slowly changing tiles.

[0051] S4. Locate and register the generated map during initialization, and assign different weights to different types of map tiles.

[0052] In this embodiment of the invention, the formula for assigning different weights to different types of map tiles is as follows: in Indicates the first i A map tile, Indicates the first i The weight of each map tile, Table of stable block weights, This indicates that the block weights are changing slowly. This indicates low dynamic block weights, and has .

[0053] The overall pose optimization objective function for localization and map registration is: in This represents the error function between the map and the observed laser (point-to-area or point-to-point distance). This indicates the current observation point cloud in the map tile. A subset within the coverage area.

[0054] S5. Calculate the rate of change of the registered map tiles and dynamically adjust the map tile type based on the mean and variance of the rate of change over multiple time windows.

[0055] Step S5 includes the following sub-steps S51 to S53: S51. Based on the registered map tiles And the current observation point cloud in the map tile subset within the coverage area The map tiles were calculated. rate of change : in Indicates the number of points or voxels occupied. The smaller the value, the smaller the change.

[0056] S52. For the same map tile, within the time window T Internal record multiple change rates And calculate the mean of the rates of change over multiple iterations. and variance ,in It represents the mathematical expectation.

[0057] S53, if If so, the map tile type will be updated to stable tiles (small mean rate of change and small fluctuations). In the case of a large average rate of change or large fluctuations, the map tile type will be updated to low-dynamic tile type; otherwise, the map tile type will be updated to slowly changing tile type. This represents the empirical threshold for the stable block mean. This represents the empirical threshold for stable block variance. This represents the empirical threshold for the low dynamic block mean. This represents the empirical threshold for low dynamic block variance.

[0058] S6. Different update methods are used to update stable blocks, slowly changing blocks, low-dynamic blocks, and ground blocks respectively, so as to achieve differentiated updates for different types of map blocks.

[0059] In this embodiment of the invention, for stable blocks, no updates are made by default; only the rate of change is counted.

[0060] For slowly changing or ground-level patches, voxel downsampling is performed on old map points to make them sparse, and then the new observation point cloud is fused in: ,in This indicates voxel downsampling.

[0061] For low-dynamic blocks, directly replace them with the latest observation point cloud: .

[0062] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for dynamic map construction and tile-based differential updates, characterized in that, Includes the following steps: S1. Divide the global map into multiple map tiles according to a fixed spatial size, and remove high dynamic targets in real time at the SLAM front end using a moving object detection method. S2. Automatically classify each map tile into a stable tile, a slowly changing tile, a low-dynamic tile, or a ground tile; S3. Generate a map based on the original laser point cloud; S4. Perform localization and map registration on the initialized map, and assign different weights to different types of map tiles; S5. Calculate the rate of change of the registered map tiles and dynamically adjust the map tile type based on the mean and variance of the rate of change over the time window. S6. Different update methods are used to update stable blocks, slowly changing blocks, low-dynamic blocks, and ground blocks respectively, so as to achieve differentiated updates for different types of map blocks; The formula for assigning different weights to different types of map tiles in step S4 is as follows: in Indicates the first i A map tile, Indicates the first i The weight of each map tile, Table of stable block weights, This indicates that the block weights are changing slowly. This indicates low dynamic block weights, and has ; The objective function for overall pose optimization in step S4, which involves localization and map registration, is: in This represents the map and observation laser error function. This indicates the current observation point cloud in the map tile. A subset within the coverage area.

2. The dynamic map construction and tile-based differential update method according to claim 1, characterized in that, Each map block obtained in step S1 independently stores point cloud data, statistical features, and dynamic state parameters.

3. The dynamic map construction and tile-based differential update method according to claim 1, characterized in that, The map tiles obtained in step S1 support on-demand loading and unloading to enable local access.

4. The dynamic map construction and tile-based differentiated update method according to claim 1, characterized in that, In step S2, the stable blocks do not change, the slowly changing blocks change continuously but slowly, the low dynamic blocks exist within a certain time scale but change significantly, and the ground blocks are obtained by the ground extraction algorithm and are processed as slowly changing blocks in the update and positioning process.

5. The dynamic map construction and tile-based differential update method according to claim 4, characterized in that, The specific steps of the ground extraction algorithm are as follows: A1. Using the location of the mobile robot as the center, convert the laser point cloud into polar coordinates, and divide the laser point cloud into partitions according to the distance, angle and radius of the laser point cloud relative to the center to obtain point cloud partitions. A2. For each point cloud partition, the group of points with the lowest position is selected as candidate ground points. SVD is used to fit the local plane, and the distance from the remaining points in the point cloud partition to the local plane is calculated. In response to the distance being less than a preset threshold, the corresponding point is divided into ground points to obtain the ground segmentation point cloud extracted in the current frame. A3. Transform the ground segmentation point cloud extracted from the previous frame into the current coordinate system, supplement the missing ground points in the current frame through multi-frame fusion, and at the same time perform a weighted average of the existing ground points in the current frame to obtain the ground block result.

6. The dynamic map construction and tile-based differentiated update method according to claim 1, characterized in that, Step S3 includes the following sub-steps: S31. Perform ground separation on the original laser point cloud to obtain ground point cloud and non-ground point cloud; S32. Store the ground point cloud directly into the ground block, and at the same time divide the non-ground point cloud into spatial blocks. S33. Set the initial type of all map tiles in the first generated map to slowly changing tiles.

7. The dynamic map construction and tile-based differential update method according to claim 1, characterized in that, Step S5 includes the following sub-steps: S51. Based on the registered map tiles And the current observation point cloud in the map tile subset within the coverage area The map tiles were calculated. rate of change : in Indicates the number of points or voxels occupied; S52. For the same map tile, within the time window T Record the rate of change multiple times and calculate the mean of the multiple rates of change. and variance ; S53, if If so, the map tile type will be updated to stable tiles. In the case of a certain condition, the map tile type will be updated to low-dynamic tile; otherwise, the map tile type will be updated to slowly changing tile. This represents the empirical threshold for the stable block mean. This represents the empirical threshold for stable block variance. This represents the empirical threshold for the low dynamic block mean. This represents the empirical threshold for low dynamic block variance.

8. The dynamic map construction and tile-based differential update method according to claim 7, characterized in that, Step S6 specifically involves: For stable blocks, no updates are performed by default; only the rate of change is counted. For slowly changing or ground-level patches, voxel downsampling is performed on old map points to make them sparse, and then the new observation point cloud is fused in: ,in Indicates voxel downsampling; For low-dynamic blocks, directly replace them with the latest observation point cloud: .

Images